Part C: Proposal Description

Multi-Agent Control:

Probabilistic reasoning, optimal co-ordination, stability analysis and controller design for intelligent hybrid systems

 

1a Research Topic

This network will focus on Multi-Agent Control. In response to the challenges presented by increased economic pressure, and the availability of low-cost computing power, a typical solution to the control engineering design task now involves the achievement of high performance at the cost of computational complexity. Such complex control systems almost always consist of a number of simpler sub-systems (or "agents") each of which addresses in a coordinated manner, a specific sub-objective or sub-task so as to attain the overall design objectives. We refer to systems of this nature as Multi-Agent Control (MAC) Systems. From a systems theory viewpoint, machines or systems constructed in this manner exhibit a rich variety of dynamic behaviour. The complexity of this behaviour arises as a result of interactions between the multiple agents and the environment in which they operate. Interactions of this kind are found in many application domains; machine-machine interaction; switched control systems and human-controller interaction. Specifically, in the context of this network, Multi-Agent Control systems are fundamental components in a wide range of safety-critical engineering systems, and are commonly found in aerospace, automotive, chemical process and power generation and distribution industries.

The design of high performance and safe Multi-Agent Control systems requires the fusion of methods from a wide range of disciplines. These include techniques from computer science, systems theory, and statistics, and their integration with established engineering design methodologies. Trained scientists with this background are simply unavailable at present. Moreover, the training of engineers with the appropriate multidisciplinary skills is not in itself sufficient to meet industry needs. Rather, there also exists a pressing need for an improved understanding of the fundamental principles underlying the dynamics of multi-agent systems. It is this lack of knowledge that has stimulated so much recent interest in the USA and Japan in such systems. The objective of this network is to service the needs of European industry, science and government by providing trained scientists, and valid methodologies, to support the design and certification of Multi-Agent Control systems. This will be achieved by the development of a principled approach to the integration of intelligent systems technologies with established engineering design methodologies for the construction of complex systems and to put these on a firm theoretical foundation.

Scientific and technological reasons for the project

Despite the prevalence of multi-agent control systems, few theoretically sound methods exist which support the analysis and design of such systems. Indeed, the design of systems of this nature is often characterised by ad hoc rules-of-thumb and extensive prototype testing. However, as systems grow more and more complex, driven by both the availability of computational power and the increasingly stringent performance requirements, exhaustive testing is rapidly becoming impractical owing to the prohibitive costs in time, manpower and finance. New, analytically-based, criteria for design, testing and safety assessment are urgently required together with a corresponding pressing need for an improved understanding of the fundamental nature of engineering systems’ embedded intelligence. In view of these considerations, the objectives of the project are not only of academic interest, but will potentially have a profound impact on current and future engineering practice and certification in Europe.

Socio-economic reasons for the project

European economic and environmental policies are concerned with improving the efficiency and competitiveness of European industry while significantly reducing pollution and, in particular, there is a requirement to significantly reduce greenhouse emissions by 2010. The network supports these objectives through the development of high-performance (non-conservative) design techniques for generic complex engineering systems, and through the application of these methods for the development of more efficient technologies for power generation and distribution (including renewable energy), petrochemical process control, aerospace and automotive design, and emissions reduction. Moreover, the network will provide a conduit for knowledge and personnel transfer between member states, and act as a vehicle for training researchers in an innovative and multidisciplinary environment. This directly supports the community objective of mobility of researchers in Europe. Scientific results from the network will be made available to the wider industrial and scientific community through, for example, workshops, summer schools, external seminars, web-based learning and incorporation of results into university curricula, and peripheral industrial collaboration.

 

1b Project Objectives

Introductory remarks

The main objective of the work is to establish the principles underlying multi-agent control systems (MACS). This will inherently require the fusion of methods from a range of disciplines including the fields of statistics, computer science, control and systems engineering.

A typical Multi-Agent feedback system is depicted in Figure 1. The multi-agent controller seeks to achieve design objectives specified by the command inputs using information from the plant and other sources. Multi-agent control systems consist of a number of dynamic elements with associated decision-making logic and information sharing paths. The decision-making logic, as a function of the internal state of the system, switches and/or schedules between the dynamic control elements. It may also instigate open-loop control actions when appropriate. Information sharing between the dynamic elements aids overall system co-ordination.

Figure 1 Structure of a typical Multi-Agent Control System

The objective of this network is the development of formal Multi-Agent Control analysis and design methods. This will be achieved through the specification and solution of a number of relevant sub-problems.

  1. To incorporate into Multi-Agent Control design, systematic decision-making techniques and, in particular, probabilistic reasoning to accommodate the inherent uncertainty in our knowledge of the state of the world.
  2. To develop tools for rigorously analysing the potentially very strong and safety-critical interaction between the outcome of decisions and the dynamic behaviour of the overall system in which a Multi-Agent Control system is embedded.
  3. To develop generic methods of control system design that support the design of the switching logic and co-ordination between multiple agents, optimisation of performance, within given constraints on the overall system behaviour.

The overall project aim will be achieved by a combination of theoretical and practical investigations to meet the following detailed objectives.

1. Development of probabilistic modelling methods for decision-support in dynamic systems

Intelligent decision making requires, either implicitly or explicitly, a model of the world that embodies both prior knowledge and measured data. New methods for integrating prior knowledge with engineering data will be investigated. The aim is the development of grey-box methods that lead to high-quality decision support systems more rapidly than was previously feasible. This will be achieved by building on powerful methods recently developed in the statistics and systems engineering fields. Specific objectives include:

  1. Extending modern non-parametric Bayesian methods to the dynamic systems context, including the development of specialised identification procedures and prediction techniques.
  2. Integrating non-parametric Bayesian representations and recently developed blended multiple model approaches to obtain a "grey box" representation supporting established divide and conquer design methods.

The methods will be used to model complex nonlinear dynamic systems, and to model human control and supervisory behaviour for the improvement of human-machine interaction.

2. Analysis of the interaction between decision outcomes and dynamic behaviour

A critical consideration in the design of complex engineering systems is their safe operation. Typically, at each level of the design hierarchy, methods exist which guarantee the safe operation of the final system. For example, methods exist for the safe specification of traditional types of engineering system, the ‘correct operation’ of software components, and the correct operation of hardware modules. However, for Multi-Agent Control systems more extensive validation is required to guarantee safety. In particular, the introduction of interacting dynamic and logic/decision-based elements will most probably introduce behaviour that is not present in any of the individual elements. Examples of such behaviour may include the emergence, in the overall system, of chaos and dynamic instability due to switching action. The aim here is the development of improved insight into the mutual interactions that exist between logic-based decision processes and dynamic behaviour, with particular emphasis on safety related aspects. Specific issues that will be investigated include:

  1. Characterisation of instability mechanisms in switched systems.
  2. Non-conservative conditions for the stability of switched linear systems. In particular, conditions based on minimum dwell time methods, Lyapunov theory and embedding theory will be studied.
  3. The relationship between information sharing (between agents) and the issue of stability

3. Design methods for Multi-Agent Control Systems

The objective is to develop systematic methods for the design of co-ordination strategies for switching and blending of multiple agents’ actions/behaviours. The behaviour and design of each individual agent is typically fairly simple and transparent. Each agent is designed with a specific sub-objective or sub-task in mind, often without taking into account inherent constraints in the system. The complexity of the system is mainly due to interactions between the multiple agents and the world when pursuing the overall objective or task of the system. The divide-and-conquer strategy has been successfully applied in this context, e.g. in gain-scheduled control where different controllers are scheduled according to the operating point of the system. Until recently, such design has been based on the assumption that each agent can be designed independently, and little support has been available for designing the co-ordinating or supervisory system. The design has been validated by simulation and typically a long range of various ad hoc modifications are necessary to make the design meets the overall specifications in the end. In order to improve the efficiency of the design cycle, we will:

  1. Integrate probabilistic approaches directly with the control design to deal with plant uncertainty.
  2. Within a probabilistic framework, develop a theory for dynamic optimisation as a design methodology for multi-agent control systems, focusing both on individual agent design and their co-ordination strategy.
  3. Develop methods for explicit characterisation of optimal multi-agent control solutions for systems with constraints on the state and input.

4. Practical application of developed method

The techniques developed will be demonstrated in practical applications including wind turbine regulation, power systems regulation and a variety of process control and aerospace applications. Each of these applications represents a challenging control problem where improved designs exploiting intelligent decision-making approaches are anticipated to lead to significant economic and environmental benefits while also providing an appropriate test-bed application for the present research.

5. Development of supporting software tools

The efficient analysis and design of dynamic systems with embedded intelligence requires the support of appropriate computer-aided design tools. It is anticipated that tools developed to support the network will be developed within the framework provided by the widely used MATLAB environment.

Potential for success

The integration of intelligent systems approaches with probabilistic reasoning and dynamic systems theory is a leading-edge area of research, which promises major breakthroughs in our ability to understand and build high-performance intelligent control technology. The powerful modelling approaches now feasible, due to new theory and modern computing power, can avoid many of the previous problems due to analytical simplifications (in e.g. noise models). Despite the challenging nature of the field, the members of the network have already made considerable progress towards addressing a number of important issues. This includes developing innovative, non-conservative analysis techniques for certain classes of Multi-agent Control systems (Shorten & Narendra 1999), establishing the feasibility of integrating blended multiple model representations with modern Bayesian non-parametric frameworks (Murray-Smith, Johansen & Shorten 1999) and deriving design guidelines for two common classes of Multi-Agent Control System (Leith & Leithead 1995,1996). Positive peer assessment of these achievements is reflected in a number of prestigious awards to members of the network, including two Marie-Curie fellowships, a Presidential fellowship and a U.K. Royal Society fellowship. This all provides encouraging evidence for the network’s probable success, and long-term benefit.

 

2. Scientific Originality

International State of the Art

Multi-agent control systems (MACS) have a rich history extending back to the earliest days of commercial computing. For example, the augmentation of standard PID controllers with max/min switches and scheduling logic has been standard practice in the chemical industry since the early 1970’s. More recently, the rapid increase in the availability of cheap computing power coupled with a cross-fertilisation of ideas from the computing and control fields has led to the current ubiquity of embedded switching/scheduling logic in the controllers for dynamic systems. Nevertheless, owing to the lack of analytic design guidelines, it has also long been known that the design of intelligent decision-making capabilities such as switching logic is often extremely complex and time-consuming for all but the simplest engineering systems. The remarkable growth in the prevalence of such MACS in recent years in many key industries of great economic importance is therefore leading to considerable interest, within both the industrial and academic communities, in the development of systematic design methods. Indeed, in many ways this is situation where practice is leading theory and driving further developments. The rapid growth of interest in the development of systematic methods is reflected, for example, in the organisation of the main US control conference, the IEEE Conference on Decision and Control. In 1998, nine specialist session were devoted to topics directly related to MACS compared to only three sessions in 1997.

A number of formal frameworks within which to consider hybrid systems have been proposed in the control literature (see, for example, Branicky et al. 1998) and these are also relevant to MACS. However, the utility of these frameworks is currently limited by the lack of fundamental understanding of the dynamic behaviour of such systems. Theoretical results relating to the dynamic behaviour of MACS are dispersed through the control, electronics and mathematics literature with many results developed in specific contexts, often independently of one another. Nevertheless, qualitative results such as the possible emergence of instability and chaos due to switching action (for example, when switching between linear systems which are individually stable) are well established (e.g. Chua et al. 1986). Current understanding of stability issues is based on long established approaches like the Circle criterion, dwell-time ideas and small-gain methods. These techniques are indirect in the sense that they establish only sufficient, but not necessary, conditions for stability and are known to be potentially very conservative. Recent work has been motivated by interest in developing less conservative methods, and has included the use of alternative indirect methods based on parameter-varying Lyapunov functions, extensions of the well know stability criteria, higher order stability criteria (quartic etc.), and Lie algebraic conditions (Branicky 94, Liberzon et al 98, Mori 98, Hespania 98, see also the survey in Leith & Leithead 1999a). Unfortunately, the degree of conservativeness of these conditions is still largely unknown and awaits the development of stability criteria which are direct, in the sense that they are based on knowledge of the possible instability mechanisms, and therefore necessary conditions.

Modelling developments

The analysis and design of multi-agent control systems and other high performance systems is almost always dependent on the availability, either implicitly or explicitly, of appropriate mathematical models of the system concerned. While modelling is a well-established field in its own right, the literature relating to model representations which specifically support control analysis and design is more recent. Many alternative mathematical representations of a dynamic system are, of course, possible. In an analysis and design context, the requirement is typically for a white- or grey-box type of model which permits a degree of physical interpretation of the model elements and provides support for established engineering design methodologies such as divide and conquer based approaches. Recently, a number of related modelling frameworks have been developed, largely independently, in the control and computing literatures, which combine classical divide and conquer (linearisation based) representations with ideas from the neural network community. These powerful new frameworks are related by their use of a blended multiple model based approach and aim to provide a strong degree of transparency combined with direct support for established engineering design methodologies (Johansen & Foss 1993, Leith & Leithead 1999b). While largely developed in the context of smooth nonlinear systems (i.e. without switched/discontinuous elements), the extension of such approaches to include multi-agent control systems seems, at least conceptually, to be relatively straightforward (Leith & Leithead 1998).

 

Probabilistic Methods

The approaches discussed so far, while powerful, are all essentially deterministic in nature with uncertainty incorporated via various types of over-bounding method. In view of the inherent uncertainty which always exists regarding the state of the world in which a system is embedded, a probabilistic framework is a natural one and in many ways this provides a viewpoint which is quite complementary to a deterministic one. Probabilistic modelling has undergone a revolution in the last 5-10 years because of the spread of computationally-intensive methods. These avoid the previously widespread analytic approximations of simplified models (e.g. conjugate families of distributions, Gaussian noise etc), and allow the scientist to develop more realistic models (see for example the BUGS software in bio-statistics). With statistics groups around the world actively pursuing such new developments, these are beginning to filter through to the computing science and control literature where there is a renewed interest in probabilistic methods. For example, note the growing convergence between the neural network and statistics literature (Jordan 1998). However, similarly to the situation regarding linear dynamic systems prior to the 1950s, the statistics literature is largely concerned with static systems and, in particular, there exist very few formal results relating to hybrid and nonlinear dynamic systems. (West & Harrison 1997) provides a modern statistical view, but unfortunately there seems as yet to be little communication between the fields.

Expected Contribution of network

The work proposed here is to move on from the state of the art outlined above. A number of significant new developments are anticipated:

  1. Probabilistic Reasoning The synergy between modern statistical methods and systems theory will be exploited to develop a new framework for probabilistic reasoning in the dynamic systems context. Modern model-based probabilistic reasoning methods provide a natural framework within which to handle uncertainty and ensure systematic decision-making. Moreover, it is clear that, in the context of analysis and design, there is potential for very considerable synergy by combining non-parametric Bayesian representations with blended multiple model representations. Specifically, this will enable the designer to work within a realistic probabilistic framework (unrealistic restrictions to analytically tractable probability distributions are avoided e.g. long tailed distributions may be used) while retaining transparency and support for established engineering design methodologies such as divide and conquer. We will also extend non-parametric Bayesian methods to dynamic systems and models of human control behaviour. There are exciting possibilities for major breakthroughs in terms of both performance and elegance of solution.
  2. Stability Existing stability results for MACS systems are either potentially very conservative or of unknown conservativeness. Similarly for performance assessments and design methods based on these results. Overly conservative designs are inefficient in the sense that they fail to realise the potential performance of a system and may consequently require an over-engineered system with associated economic implications. The development of non-conservative (or at least with greatly reduced conservativeness which is well characterised) stability tests for a useful class of systems would represent a very significant breakthrough of great practical importance.
  3. Design Techniques based upon for multi-agent control systems will be developed. This work extends currently existing methods that have been developed single agent systems. The work on constraint handling will focus on developing explicit controllers rather than implicit controllers that require extensive online optimisation as in for example Model Predictive Control. Methods which integrate the probabilistic modelling approaches directly with control design will be investigated.

The overall research contribution of the network will be the integration of intelligent systems approaches with probabilistic reasoning and dynamic systems theory. This involves bringing together a number of areas in an innovative and rigorous manner, leading to major breakthroughs in our ability to understand and build high-performance intelligent control technology. The efficacy of the new design methodology will be established with respect to a number of challenging practical applications.

 

3. Research method

Probabilistic modelling methods for decision-support in dynamic systems

The objective is the development of new "grey-box" modelling approaches which, firstly, provide a systematic framework for integrating prior knowledge with engineering data while, secondly, retain strong support for well-established engineering design methods such as divide and conquer. With regard to the first requirement, Bayesian methods provide a very general and systematic framework for integrating prior knowledge with measured data. Specifically, modern non-parametric Bayesian methods offer many advantages including the retention of direct contact with the measured data, good conditioning in high-dimensional spaces even when data is limited, and support efficient adaptation to permit the ready exploitation of new data as it becomes available. However, non-parametric Bayesian representations have currently only really been investigated in the context of static (rather than dynamical) systems, are essentially "black box" in nature and at present provide little direct support for divide and conquer design. With regard to the second requirement, recently developed velocity-based blended multiple model methods provide a powerful "grey box" framework which directly supports divide and conquer design for dynamic systems. However, velocity-based approaches presently assume the prior existence of a suitable model obtained, for example, from first principles. Clearly, the potential for considerable synergy exists from combining non-parametric Bayesian methods with the velocity-based blended multiple model framework. The research will proceed as follows.

M.1 Extend non-parametric Bayesian methods to dynamic systems. Initially, methods for Gaussian process priors will be investigated since this is a tractable yet still very powerful class of non-parametric representations. The extension to dynamic systems will require

a) Specialised identification methods.

b) Prediction methods (model output is a probability distribution, rather than a single value, and this needs to be propagated through the system dynamics in order to determine future predictions).

M.2 Integrate non-parametric Bayesian methods with the velocity-based "grey box" framework. In particular, study velocity-based methods for establishing the nonlinear structure of the non-parametric model.

M.3 Extend consideration to more sophisticated non-parametric models. Models with non-Gaussian noise will be investigated.

M.4 Extend consideration to mode-switching systems. Markov mixtures methods, as used in e.g. (Murray-Smith, 1998), will be studied with particular application to the modelling of both multi-agent systems and human operators.

Analysis of the interaction between decision outcomes and dynamic behaviour

The objective is the development of non-conservative analysis methods for determining the stability of multi-agent systems. Despite the occurrence of multi-agent control systems in many application areas, the lack of dynamic stability criteria will prohibit their success in safety critical applications such as those found in the automotive and aerospace industries. The absence of satisfactory stability criteria can be attributed to a number of factors. Problems in this area are, academically, highly challenging, and tools to develop insights into multi-agent systems, i.e. cheap powerful computers, have only recently become available. Moreover, the availability of high-speed computers has encouraged a trend away from analysis of dynamic instability, toward ‘search and simulate’ strategies for the solution of (purported) NP-hard problems. This research will build upon new results recently obtained by Dr. Shorten and Prof. Narendra at Yale University. The approach adopted is to develop fundamental insights into instability mechanisms in switching systems, and to express these in the form of Aizermann-like (Slotine & Li 91), or matrix pencil conditions. This strategy has led to new fundamental quadratic stability criteria for classes of multi-agent control systems. Extensions of this work is likely to lead to new non-quadratic criteria, as well as novel dwell-time criteria for the stability of a wide class of multi-agent control system. The specific issues to be addressed are:

A.1 Characterisation of instability mechanisms: The principal tools used will be geometric state space tools, and the results will be expressed in the form of matrix pencils.

A.2 Non-conservative conditions for the stability of switched linear systems: As in A.1, the principal tools adopted will be geometric state space tools. The principal analysis tools used here will be the common Lyapunov function (not necessarily quadratic), and eigenstructure analysis for dwell time determination.

A.3 Information sharing (between decision making agents) and stability: The methodology utilised here will exploit the structure of the application under consideration. It is anticipated that rank-1 eigensystem perturbation results, and results from linear algebra concerning simultaneous triangularisation of matrices, will be of use here.

Design methods for MACS

The objective is to develop systematic methods for the design of co-ordination strategies for switching and blending of multiple agents’ actions/behaviours. Conventionally, the design of a multiple-agent control system is typically validated by simulation and involves an iterative ad hoc approach which is extremely time-consuming for all but the simplest systems. There is, therefore, a strong incentive to improve the efficiency of the design cycle. Design tools based on the probabilistic reasoning and stability analysis methods developed by the other strands of the project will be investigated. In particular, the probabilistic modelling approaches developed by the network will be used to provide a very natural framework within which to formulate the requirements associated with robust design. Issues to be investigated include

D.1 Robust analysis.

a) Ensuring the robustness of the uncertainty description itself to deficiencies in prior knowledge and measured data.

b) Markov-Chain Monte Carlo methods for evaluating robust stability and performance.

D.2 Optimisation-based methods. Within a probabilistic framework, dynamic optimisation will be studied as a design methodology for multi-agent control systems focusing both on individual agent design and their co-ordination strategy. In particular,

a) Methods for mixed-integer optimisation and the branch-and-bound technique will be investigated for designing multi-agent systems involving both discrete and continuous variables.

b) Design methods based on the Monte Carlo analysis methods developed in D.1b will be investigated.

D.3 Constrained systems. Real systems are always subject to various physical constraints. In high performance systems, the accommodation of these constraints typically leads to the requirement for a multi-agent type of control solution. Multi-agent design methods for constrained systems will be specifically studied. It is anticipated that dynamic programming methods formulated in terms of Linear Matrix Inequalities (LMIs) will prove valuable and investigations will include the explicit characterisation of optimal multi-agent control solutions for systems with constraints on the state and input.

Practical application of developed methods

In view of the challenging nature of the problems being studied, it is anticipated that a combination of complementary "top-down" and "bottom-up" investigations will greatly facilitate progress. Real-world applications will, in addition to being used to demonstrate general methods which have been developed, therefore also be used on an on-going basis as test-beds with the intention that practical solutions developed in response to realistic problems will help stimulate general theoretical developments in a bottom-up manner. Benchmark applications will be developed in the aerospace, power generation and process control fields in close collaboration with BAe, CERPD and INEA. The industrial partners will play a key role in the definition of the problems considered, including the contribution of data where appropriate, and will assist in assessing the outcome of the investigations. These industry-driven applications will be supported by detailed experimental studies using wind turbine and NOx separator test-rigs. Both are examples of highly nonlinear dynamic systems which present challenging control problems. These are of considerable interest in their own right with improved control designs exploiting intelligent decision-making approaches are anticipated to lead to significant economic benefits. The research will include:

P.1 Industry-driven applications studies

P.2 Experimental investigations.

a) Wind turbine regulation.

b) NOx separator regulation.

These investigations will involve (i) details of the test-rigs (including validated simulation models) will be distributed to the network, (ii) application of developed modelling methods, (iii) implementation issues associated with multiple-agent controllers, (iv) investigation of multiple-agent control strategies, (v) detailed performance assessment and comparison with existing control designs. Data and results will be made available to the network members.

We believe that this healthy mix of leading-edge theoretical developments, and practical application to real processes is a promising way of achieving significant and useful progress.

4. Workplan

Distribution of Tasks

For the purposes of co-ordinating research work, the network partners will initially be organised into the following (overlapping) results-oriented teams. These have been split up by speciality, but the intention behind the network is to stimulate exciting new developments in theory and practice, by bringing together experts within the network and by the introduction of knowledge from invited external experts. The named teams and topics are intended to act as a focus to the project participants in order to make concrete progress, but we expect significant contributions in new directions not outlined in detail. An important aspect of this is the collective "Roadmap" document we shall begin working on from the start, which as well as providing a resource for other researchers, should be an ideal way to kick-start the network.

Team Co-ordinator

Team Members

Probabilistic Modelling Team

GU

GU, SU, DTU

Dynamic Analysis Team

NUIM

NUIM, SU

Design Methods Team

NUST

NUST, DTU, GU, IJS

Applications Team

IJS

SU, IJS

Software Development

NUST/GU

NUST, GU, SU, DTU, NUIM, IJS

Schedule

The network will run for 48 months, but the maximum contract length for the young researchers will be 40 months. This provides the network co-ordinators with the flexibility needed to acquire top-quality staff. The overall network schedule to be followed by the research teams is outlined in the following table:

Project Milestones

Topic

To be achieved by mid-term review:

To be achieved by final review:

Probabilistic Modelling for Decision Support in MACS

  1. Literature review of probabilistic modelling and modelling methods for analysis/design (to feed into "road-map" report)
  2. Identification methods for non-parametric Gaussian process priors in dynamic systems context investigated.
  3. Prediction methods for non-parametric Bayesian models of dynamic systems investigated.
  4. Investigation of methods to accommodate long-tailed/robust noise distributions

  1. Integration of Gaussian process prior representation with blended multiple model framework
  2. Application of non-Gaussian noise assumptions in dynamic system modelling and control
  3. Human operator modelling applied in simulation environment.

Analysis of MACS

  1. Literature review of stability results (to feed into "road-map" report).
  2. Characterisation of stability mechanisms in a wide class of MACS.
  3. Investigation of quadratic and non-quadratic criteria MACS stability.

  1. Less conservative dwell-time results for MACS investigated.
  2. Local-state stability criteria studied for a wide class of MACS.

Design Tools

  1. Literature review of design methods (to feed into "road-map" report).
  2. Integrating probabilistic models/Monte Carlo methods with control analysis & design studied.
  3. Development of tools for design of multi-agent systems using dynamic optimisation

  1. Development of a dynamic programming solution to the constrained optimal multi-agent control problem
  2. Case study using software design tool prototypes incorporating the developed methods

Practical Applications

  1. Initial wind turbine benchmark established for use by network
  2. Establish benchmark control system for selected process from semi-industrial laboratory.
  3. Demonstration of probabilistic modelling methods on wind turbine benchmark.
  4. Experimental evaluation of probabilistic modelling methods on semi-industrial laboratory’s processes.

  1. Investigation of MACS control strategies for wind turbine regulation.
  2. Detailed performance assessment/ comparison of MACS with existing wind turbine control methods.
  3. Investigation of MACS for regulation of NOx separator.
  4. Detailed performance assessment and comparison with existing process plant control designs.

Training/

Dissemination

  1. Initial induction workshop
  2. Web site established.
  3. Second workshop to induce late starters (attached to first review meeting)
  4. 1st Summer school at IJS, Ljubljana.
  5. Publication of "road-map" review paper on MACS.
  6. Internet publication of initial prototype software developed for modelling, analysis and design of MACS.

  1. Workshop affiliated (subject to organisers consent) to European Control Conference 2001
  2. 2nd Summer school at NUIM.
  3. Workshop affiliated with major statistics conference.
  4. Workshop affiliated (subject to organisers consent) to Conference on Decision & Control, 2002
  5. Dissemination of results via academic publications, internet publishing & direct contact with interested parties
  6. Internet publication of developed software.

Reporting

Mid-term report to EC

Final report to EC

Each young researcher will also report on their personal contribution on completing contract.

5. Collective Expertise

Introductory Remarks & Complementarity of Expertise

The primary participants consist of five universities and an industrial research centre, in five European countries: Ireland, UK, Denmark, Norway and Slovenia as detailed below. These core members of the network will be augmented and supported by a substantial group of industrial organisations closely affiliated to the network which are also detailed below. The members and affiliates of the proposed network all have long-standing interests in the analysis and design of MACS. This common interest has led naturally to their on-going collaboration. As discussed in part 1, the study of MACS requires a fusion of skills from a number of scientific disciplines (including computer science, statistics, systems and control engineering) and this is reflected in complementary yet tightly integrated skills of the network members, detailed below and summarised in the following table.

Partner

Discipline

Primary Research Speciality

GU

Comp. Sci., Systems & Control

Probabilistic reasoning, Modelling for analysis/design

SU

Elec. Eng.

Analysis, design & modelling of nonlinear/switched systems

NUST

Eng. Cybernetics

Control design tools, Modelling for analysis/design

DTU

Statistical Modelling

Statistical modelling & Forecasting

NUIM

Comp. Sci., Maths

Analysis of switched systems, Modelling for analysis/design

IJS

Automation & Control/Industrial

Application of advanced control design methods

Affiliated Organisations

British Aerospace (BAe). British Aerospace is one of Europe’s leading defence and aerospace groups employing over 43,000 staff and with annual sales exceed EUR 12 billion. In the commercial aircraft market, British Aerospace is a partner in the leading European consortium Airbus Industrie (together with Aerospatiale, ,Daimler-Benz Aerospace and CASA) and builds a range of regional jet aircraft. BAe and the Universities of Strathclyde and Glasgow have been collaborating closely for longer than 7 years. The work, which has been funded by the U.K. government and privately by BAe itself, is concerned with the design of the complete aircraft system. Being an amalgamation of a large number of sub-systems (such as the flight control system, the engine control system, the environmental control system) which all interact, the aircraft system constitutes a multi-agent system.

CERPD. The Centre for Economic Renewable Power Development (CERPD) is a recently established joint venture between the Universities of Strathclyde and Glasgow which acts to provide a focus for multi-disciplinary research on power systems generally and renewable energy systems in particular. The CERPD has a strong economic and industrial emphasis with members including the engineering schools of the Universities of Strathclyde and Glasgow and the Strathclyde Graduate Business School. Supporting companies include ABB, Motorola, Rolls Royce, Westinghouse and many of the companies involved in power production and distribution in the U.K. The Centre also has a strong interest in renewable energy aspects of power generation with close links to companies including the wind turbine manufacturer and operator Renewable Energy Systems and the specialist European consultancy firm Garrad Hassan & Partners. The network has very close links to the CERPD. In particular, Dr. W.E. Leithead of SU who is a founding member of the CERPD.

INEA. INEA was initially established over ten years ago by the Institute Jožef Stefan in 1987 with the objective of accelerating technology transfer to industry. As an independent private company, INEA has grown rapidly (currently employs around 50 engineers) as a control system vendor and high technology solution provider specialising in the areas of industrial energy management and the control of industrial processes.

Other industrial organisations closely affiliated to the network include SINTEF (an independent industrial research organization in Norway, the largest independent research consultancy in Europe, employing around 1400 engineers and with close links to NUST and DaimlerChrysler) and MAN B&W (a specialist Danish engineering firm involved in engine applications with close links DTU).

University of Glasgow (GU)

The role of Glasgow.

The University of Glasgow will provide the hub of the network, having research interests and personal contacts with all other nodes. The work at GU will concentrate on development of statistical approaches and intelligent decision systems which can aid modelling and control work.

Key personnel at Glasgow.

Name

Affiliation

Time

Roderick Murray-Smith

Department of Computing Science

20%

D. M. Titterington

Department of Statistics

3%

Expertise and competence of Glasgow

The Department of Computing Science is one of the best in Britain, recently given a Queens Award for its outreach policies to industry and commerce, was awarded the highest grading (5*) in the UK Research Assessment Exercise, and also given the highest rating for teaching. Dr. Murray-Smith has worked in industrial research at Daimler-Benz in Germany for 7 years, as well as holding academic positions at M.I.T. and the Technical University of Denmark (as a Marie Curie Fellow) before coming to GU. The contribution from GU is strengthened by excellent neighbouring departments. As noted in this proposal, statistical tools will play an increasingly important role in intelligent systems. Ongoing collaboration with Prof. D. M. Titterington in Statistics will provide significant expertise to the network from a leading statistician, and there will be interaction with his colleague Dr. A. Nobile, who has an international reputation in Bayesian statistics. Within the Centre for Systems and Control (of which R. Murray-Smith is also a member) Prof. K. Hunt’s (an ex-colleague from DaimlerBenz) research group of approximately 10 doctoral students and post-docs is also involved in leading-edge research in modelling and control designwhich broadens the scientific base in Glasgow. The human modelling work benefit from virtual reality equipment provided by Prof. M. Atkinson, and the well-funded Revelation project.

Research linkage and existing connections between collaborators

Glasgow and Strathclyde Universities are "preferred partners" - the university administrations have agreed to close collaboration between these two physically close (3km apart) universities. The Centre for Economic Renewable Power Development is run jointly by both universities. R. Murray-Smith has worked with most members for some time. This is documented in joint publications (co-edited book cited below, and co-authored a number of scientific papers with R. Shorten and T. A. Johansen.) Met other members in a number of workshops/conferences (all were invited speakers at workshops in Trondheim, Copenhagen, and a coming conference in Karlsruhe). In terms of formal collaboration, R. Murray-Smith while a project manager at Daimler-Benz Research, managed a formal research project between SINTEF, Daimler-Benz research and GU. He worked in the same research group as Dr. Shorten at Daimler-Benz for four years, and worked in the same group as Dr. Rasmussen in the Technical University of Denmark for 2 years, running courses and conference workshops together. Dr. Leith was recently on a study visit to Copenhagen to work with R. Murray-Smith.

Two most-relevant publications

 

University of Strathclyde (SU)

The role of University of Strathclyde.

The work at SU will concentrate on the modelling and analysis of MACS. In addition, SU will act to co-ordinate the wind turbine test-bed application including the assessment, in the context of renewable energy systems, of the economic and technological benefits of the methods developed by the network.

Key personnel at University of Strathclyde.

Name

Affiliation

Time

Douglas Leith

Department of Electrical Engineering

20%

W.E.Leithead

Department of Electrical Engineering, Centre for Economic Renewable Power Development

10%

Expertise and competence of University of Strathclyde

The Department of Electronic and Electrical Engineering at the University of Strathclyde is a well-established international centre of excellence in both control engineering and renewable energy systems with extremely strong industrial links. Dr. W.E. (Bill) Leithead is Reader in Control Engineering and a founding member of the CERPD. In addition to control design for nonlinear systems, Dr. Leithead’s research interests include wind turbine dynamics and control, multivariable control and implementation aspects of control design. With regard to wind turbine modelling and regulation, Dr. Leithead has led more than twenty projects funded by the U.K. Engineering and Physical Sciences Research Council, the Department of Trade and Industry and directly by industry. Dr. Leithead also has strong links with the aerospace industry and is actively involved in the development of design methodologies for multivariable control applications. Dr. Douglas Leith graduated in 1986 with a first class joint degree in Electrical Engineering and Computer Science. During this course of study he was awarded prizes including the Alexander J. Younger Memorial, the John Oliphant Bursary, the Sir John Pender Bursary, the ICI Prize in Control Engineering and the Howe Prize. Following PhD studies on stochastic experiment design, Dr. Leith has pursued research interests in the analysis, design and modelling of nonlinear and switched-linear systems with over 50 academic publications in this area. This work is directly related to the research proposed in this document. Positive peer assessment is reflected in the award (in competition with applicants from all areas of science and engineering) to Dr. Leith in 1995 of a prestigious ten-year Royal Society personal research fellowship to support this research.

Research linkage and existing connections between collaborators

Dr. Leith and Dr. Leithead have a long history (over 7 years) of close collaboration, including holding a number of joint research awards, which has been documented in numerous academic publications. There is also a significant history of both formal and informal collaboration with all of the other partners of network. In addition to technical exchanges, informal visits and discussions at international conferences, this includes the award of a British Council research grant jointly to Dr. Shorten and Dr. Leith, funded study visits to Dublin and Copenhagen to pursue joint research with, respectively, Drs. Shorten and Murray-Smith, joint involvement in workshops held in Trondheim, Copenhagen and a special session to be held at the European Control Conference in August 1999.

Two most-relevant publications

 

National University of Ireland, Maynooth (NUIM)

The role of NUI, Maynooth.

NUI-Maynooth will consider performance and stability aspects of a class of MAC systems. This task will be carried out as a collaborative effort between the Computer Science and Mathematics Department and NUI Maynooth, and with the Department of Electronic Engineering at NUI, Dublin.

Key personnel at NUI, Maynooth.

Name

Affiliation

Time

Robert Shorten

Department of Computer Science, NUI, Maynooth

20%

Tom Dowling

Department of Computer Science, NUI, Maynooth

5%

Fiacre O’Cairbre

Department of Mathematics, NUI, Maynooth

5%

Paul Curran

Department of Electrical Engineering, NUI, Dublin

10%

Expertise and competence of NUI, Maynooth

The analysis and design of signal processing systems is one of the main research strengths of the Department of Computer Science at NUI, Maynooth. Some of this work has been supported by the EU research projects; see ESPRIT research project 21894; ESPRIT research project 23009; ESPRIT research project 20557; TMR contract number ERBFMBICT950092 and HCM contract number ERBCHBICT930711. Within this framework, Dr. Shorten has been pursuing his primary research interest, namely that of the stability of hybrid and time-varying dynamical systems. This work is directly related to the research proposed in this document. Dr. Shorten has been the recipient of several awards to pursue these activities. From 1993 to 1996 he was the holder of a Marie Curie Fellowship at the Daimler-Benz Research Institute in Berlin. The results of this research were presented in the form of a Ph.D. thesis, academic publications, and as an industrial patent with Daimler-Benz. His work has also been profiled in an EU research document, ‘Marie Curie Fellowships –Success Stories’, and his Ph.D. thesis was nominated in part for the Mercedes Benz research award. In 1996 Dr. Shorten was invited to continue this work as a visiting post-doctoral fellow at the Center for Systems Science, Yale University with Professor K. S. Narendra. In 1997 he was awarded a European Presidential Fellowship to return to Ireland to work in the area of stability theory. Full details of publications related to this research, plenary lectures on this topic, as-well as other bibliographical details, can be found at www.cs.may.ie/~rshorten.html. These activities are complemented by the participation of Dr. Curran in the network. He is also a recipient of several research awards for his work in stability of non-linear dynamic systems, including a prestigious Fulbright award, which he received in 1996. The Mathematics department at NUI, Maynooth is one of the leading Mathematics departments in Ireland. The study and analysis of dynamic systems is an integral component of departmental teaching and research programmes. Furthermore, the department possesses considerable expertise in the area of functional analysis. This expertise is of particular relevance for the stability analysis proposed in part 1b.

Research linkage and existing connections between collaborators

Dr. Shorten has a history of colloboration with the above scientists and other partners in the network. His work with Dr. Curran, Dr. Murray-Smith and Dr. Johansen has been documented in several peer reviewed publications, and his work with DaimlerChrysler has resulted in an industrial patent. Previous informal research collaboration with Dr. Leith has been recently formalised in British Council research grant. Dr. Murray-Smith and Dr. Shorten have also collaborated for over 4 years in an industrial context while employed at DaimlerChrysler, as well as with Dr. Johansen who is a consultant with DaimlerChrysler.

Two most-relevant publications

 

Norwegian University of Science and Technology (NUST)

The role of NUST.

The work carried out at NUST will focus on optimization based methods for the design, analysis and implementation of multi agent control systems. The objective is to develop coordination strategies for switching and blending of multiple agents actions that minimize suitable optimization criteria while satifying constraints on their collaborative behaviour and performance.

Key personnel at NUST.

Name

Affiliation

Time

Tor A. Johansen

Department of Engineering Cybernetics

15%

Bjarne A Foss

Department of Engineering Cybernetics

5%

Expertise and competence of NUST

Profs. Tor A Johansen and Bjarne A. Foss have been working on multi-models and multi-controllers and published 15 papers in scientific journals since 1992 in this area alone, in addition to several papers in edited volumes and conferences as well as several papers in other areas of control and systems engineering. Prof. Tor A Johansen was a graduate student between 1991-1994. Between 1995-1997 he was a researcher with SINTEF, and since 1997 an associate professor at NUST. He has been involved in several industrial projects as a researcher and consultant. He is currently an associate editor of Automatica and IEEE Trans. Fuzzy Systems. Prof. Bjarne A. Foss became a professor at NUST in 1990. Prior to that he was a group leader at SINTEF. He has wide experience in research leadership from both industry and academia.

The Department of Engineering Cybernetics employs 12 professors and 7 adjunct professors covering a wide range of areas such as control engineering, industrial computer systems, robotics, process control, optimisation, real-time systems, vehicle control etc. The Department of Engineering Cybernetics has a very strong PhD programme with approximately 5-10 PhD students graduating each year. Several current projects focus on topics, including hybrid control, optimization based control and multi-controller systems, which are closely related to the work in the present proposal. These are funded mainly by the Research Council of Norway. NUST has close collaboration with the industrial research organisation SINTEF, which employs about 1400 researchers, and shares local facilities and personnel. SINTEF is, in addition to being an independent organization, an ideal route for dissemination of results to industry. The ORBIT software has been developed as a joint venture between NUST and SINTEF. It provides a tool for development and analysis of multi-models and multi-controllers. The ESPRIT LTR project H2C, where SINTEF is a partner, can complement certain aspects of this proposal.

Research linkage and existing connections between collaborators

Tor A Johansen has been collaborating with Roderick Murray-Smith and Robert Shorten informally and through joint industrial projects with DaimlerChrysler and Glasgow University since about 1993. The focus of his work has been on multi-models and multi-controllers and resulted in several joint publications, including one jointly edited book. He has also organised and participated in joint workshops with other network partners in Trondheim, Copenhagen, and soon Karlsruhe.

Two most-relevant publications

 

Technical University of Denmark (DTU)

The role of DTU

The Department of Mathematical Modelling (IMM) at the Technical University of Denmark (DTU) is well suited to strengthening the statistical background in the network. DTU will be involved in training other nodes in the use and implementation of modern statistical simulation software, and will actively investigate the overlap between computationally-intensive statistics and dynamic systems theory.

Key personnel at DTU

Name

Affiliation

Time

Carl Edward Rasmussen

Department of Mathematical Modelling

15%

Lars Kai Hansen

Department of Mathematical Modelling

5%

Expertise and competence of DTU

Dr. Rasmussen has extensive experience with implementation of Markov Chain Monte Carlo methods for Bayesian inference in intelligent systems environments. He also has contributed to leading-edge research in Gaussian Process priors for non-parametric Bayesian models. Current research topics include analysis of complex (non-Gaussian) noise models and Monte Carlo based inference in complex models trained on small data sets, both topics that are of immediate relevance to the design of controllers. Dr. Hansen has 10 years experience with modelling of complex tasks, and has published extensively on neural network ensembles, regularisation and selection. He has supervised 15 PhD students.

The group has made extensive contributions to model estimation, regularisation and selection in the traditional approaches as well as practical experience with Bayesian non-parametric modelling based on modern Monte Carlo techniques for integration in high-dimensional spaces. The enormous growth in computational capacity of desk-top computers, has made these methods feasible for practical modelling problems. These techniques have been taught at the PhD level course in Advanced methods for non-parametric Modelling (in 1999 with Dr. Murray-Smith), which strongly emphasises the theoretical foundations as well as familiarising the participants with leading edge software, from both IMM and elsewhere.

They are backed up by the extensive skills-base of one of Europe’s largest mathematical modelling departments, with over 50 employees, providing useful sources of information, contacts and training modules for other nodes in the network. Both are in the Digital Signal Processing section of the institute, a lively group of around 15 Post-graduate researchers. The group has very active research links to a number of medical research institutes, and information technology companies, who often call upon the staff as consultants. Ramboll Informatics, for example, has used modelling skills from the group similar to those to be used in the research network. The section has close collaboration with MAN B&W (acoustic and chemical lubricant analysis for Diesel engine monitoring) and is participating in the international EU BIOMED2 and the Human Brain Projects. The "Lyngby" Matlab Brain Imaging toolbox designed there is used at Harvard University Department of Psychology and the Center for Magnetic Resonance Research, Minneapolis.

Research linkage and existing connections between collaborators

R. Murray-Smith (now at GU) worked in the same group with C. Rasmussen at DTU for 2 years and both have run courses and conference workshops together. Other members of the network were at DTU for a workshop in April 1999, to give everyone a chance to meet informally before submitting the proposal.

Two most-relevant publications

 

Institut Jožef Stefan (IJS)

The role of IJS, Ljubljana

The work at the IJS will concentrate on the implementation aspects and "bottom-up" investigation of MACS whereby the insight provided by a variety of practical applications will be exploited to provide a focus for theoretical developments.

Key personnel at IJS

Name

Affiliation

Time

Juš Kocijan

Department of Computer Automation and Control

15%

Damir Vrani

Department of Computer Automation and Control

5%

Expertise and competence of IJS

IJS is the most prominent research institute in Slovenia. It incorporates departments from various fields of science from physics, chemistry to computer science. All departments at IJS are involved in numerous international projects and have contacts on all continents. Beside this fact the policy of IJS is to provide constant inflow of up-to-date technology to industrial practice what, consequently, results in great number of industrial applications, patents and technology improvements. The Department of Computer Automation and Control is engaged in a broad range of research, development and applications in the area of computer-based control of engineering systems, together with the provision of associated services, engineering consultancy and education. The activities are split over four major areas of work:

The department is involved in international projects like TEMPUS, COPERNICUS, COST and bilateral projects with Germany, Italy and the Czech republic.

The department supplements its body of expertise within a virtual organisation, the so-called Technology Vertical which consists, beside the department, of the INEA company (SME) and two laboratories from the Faculty of Electrical Engineering, University of Ljubljana.

Research linkage and existing connections between collaborators

There has already been fruitful contact between the University of Glasgow and the Jožef Stefan Institute and University of Ljubljana in recent years through an EU-funded TEMPUS programme. The aims and objectives of this programme were essentially educational but it did also provide an opportunity for technical exchanges and preliminary discussions about future research co-operation. The University of Strathclyde was also included in research activities that came out of this link. The TEMPUS connection led on to a British Council-funded ALIS link which has provided more opportunities for contact and for discussions between researchers in Ljubljana and in Glasgow about areas of common research interest. In particular, Dr. Kocijan has visited both the University of Glasgow and the University of Strathclyde five times since 1993, including a 3-month secondment in 1993 and a 2-week study visit in 1999. A number of joint publications have resulted from these contacts, most of these concerned with the results of various theoretical and application studies.

Two most-relevant publications

 

6 Collaboration

The research outlined in this document is motivated by a clear industrial objective; the development of systematic procedures for the design of Multi-Agent Control systems. The members of the proposed network all have long-standing interests in the analysis and design of Multi-Agent Control systems. This common interest has led naturally to their on-going collaboration in an informal network which the current proposal would both put on a formal basis and enable funding for transnational research in a direct manner which is difficult to achieve through purely national funding mechanisms. As discussed in part 1, the study of Multi-Agent Control systems requires a fusion of skills from a number of scientific disciplines (including computer science, statistics, systems and control engineering) and this is reflected in complementary yet tightly integrated skills of the network members, see table 6.1.

Partner

Discipline

Primary Research Speciality

GU

Computing Sci., Systems & Control

Probabilistic reasoning, Modelling for analysis/design

SU

Electrical Engineering

Analysis, design & modelling of nonlinear/switched systems

NUST

Engineering Cybernetics

Control design tools, Modelling for analysis/design

DTU

Statistical Modelling

Statistical modelling & Forecasting

NUIM

Computing Science, Maths

Analysis of switched systems, Modelling for analysis/design

IJS

Automation & Control/Industrial

Application of advanced control design methods

Table 6.1 Primary skills of network constituents

The results of the long history of successful collaboration between the network members have been documented in the form of peer reviewed academic publications, joint research grants, patents, and in the form of industrial applications. Evidence of the success of these activities is visualised in the map below, which shows links in the network which have already had deep collaboration i.e. worked together physically, published joint scientific articles, had joint formal projects. Specifically, this history of collaboration includes (also see section 5):

This strong history of collaboration, the specification of joint project objectives, lack of weaker nodes, and the overlap in scientific interests will ensure successful working-level interaction. Knowledge transfer between institutions is central to the ethos of this network, both in terms of research training and research activity. This will aid and encourage global interaction between network nodes and cross fertilisation of ideas. At a practical level, the network partners will interact at a day-to-day level via modern electronic communications (mail, discussion groups, telephone) and extended personnel transfers. Progress will be monitored and co-ordinated through regular project meetings (both electronic and physical meetings). It is important to emphasise that the project partners have a substantial history of successful working-level collaboration, using the measures described above, and that regular meetings already form an integral part of their working relationship. Trans-national funding of network activities will enable significant progress to be made, and sustain a ‘critical mass’ of research which would be difficult to achieve within the constraints of purely national funding mechanisms.

 

7. Organisation and management

The network is a tightly integrated one with a significant history of previous and current collaboration between the partners. In particular, all of the partners are experienced in working together remotely over the internet (successfully executing a number of formal research projects at distance, as well as editing a book, and writing journal papers). The Universities of Glasgow and Strathclyde will form the hub of the network with the University of Glasgow acting as formal network co-ordinator.

Co-ordination of Reporting, Research and Training

Reporting GU is overall co-ordinator, responsible for collating results to form mid-term and final reports to EC. Nodes are nevertheless individually responsible for timely production of reports on their work for progress meetings and to meet EC review requirements. GU will chair network meetings which will be held approximately every six months to review and document progress, particularly with regard to the training received by the young scientists.

Co-ordination of Research The research objectives require the close collaboration and effective co-ordination of the network partners. For the purposes of co-ordinating research work, the network partners will initially be organised into the following results-oriented teams. This overlapping team organisation is intended to be flexible in nature with, for example, a degree of natural re-organisation anticipated as the project progresses.

Team Co-ordinator

Team Members

Probabilistic Modelling Team

GU

GU, SU, DTU

Dynamic Analysis Team

NUIM

NUIM, SU

Design Methods Team

NUST

NUST, DTU, GU

Applications Team

SU/IJS

SU, IJS

Software Development

NUST/GU

NUST, GU, SU, DTU, NUIM, IJS

Co-ordination of Training Each network partner is responsible for local training. The network-wide aspects of the training programme will be co-ordinated by GU who will delegate responsibility for organising specific activities (e.g. summer schools) to the other nodes as appropriate. An initial familiarisation week is planned to allow young researchers to meet and do some common training to "warm-up" the network.

Organisational Competence

Previous experience of Dr. Murray-Smith, GU, as network co-ordinator includes running a number of large international, multi-centre research projects, with direct supervision of PhD-qualified research scientists. These involved both academia and industry in Europe and the USA, while he was employed as a project manager at DaimlerBenz research in Berlin, Germany. This includes several industrial research projects over 2-3 years involving DaimlerBenz, DASA Satellites division, and Eurocopter, an international collaboration with M.I.T. in the USA, and a long-term research project involving SINTEF in Norway, and GU in Scotland. The Department of Computing Science at GU also has extensive experience of administering EC-funded research networks such as the MIRO network in Information Retrieval in the 4th Framework programme. The coordinators at the other nodes have similarly rich experience, and given our past collaboration experience, we are confident of success.

Dissemination of Results

Project results will be actively disseminated through a number of channels including

  1. Publication of academic papers and presentations to relevant international conferences.
  2. Organisation of specialist workshops affiliated to international conferences.
  3. Running of summer schools open to, and actively encouraging, participants and invited experts from outwith the network.
  4. Development of demonstration software.
  5. Publishing pre-prints of papers, scientific reports, demonstration software, announcements for and material from workshops/summer schools on the internet. GU will maintain this web-site.

In addition, results will be disseminated by direct contact with interested parties including network industrial affiliates British Aerospace, INEA, CERPD, SINTEF and DaimlerChrysler.

8. Training need

Training need in area covered by project

The availability of low cost computing has fundamentally altered the nature of engineering practice. A typical engineering solution now involves achieving high performance through computational complexity. Such performance and computation- driven problem solving techniques require graduates with a novel set of engineering skills. These include competence in systems and control; probability and statistics (decision theory, modelling) and software engineering. However, traditional engineering and computer science programmes have been slow to respond to the changing needs of industry. Typically, engineering graduates may be strong in systems and control theory, but weak in decision theory and software skills. Similarly, computer science graduates are strong in software skills, and perhaps decision theory, but weak in systems and control. The existence of a significant skills gap is clear both from the strong support for the network objectives expressed by major high-tech European organisations such as British Aerospace, INEA, CERPD, SINTEF and also from the direct experience of the network partners while in industry. We cite recent policy documents in the references, which support this viewpoint. The development and validation of real engineering systems was often delayed due to a lack of availability of practical design tools and a lack of theoretical insight into the deeper technical problems. A major objective of this network is to address the needs of industry and government through the creation of a network of trained scientists with the required multi-disciplinary skills.

Figure 8.1 Structure of network

Enhancement of European Human Potential

The proposed network satisfies the objective of enhancing European human potential in a number of ways. Primarily, the network will enhance Europe’s industrial and scientific competitiveness, in the problem domain addressed by the project, through the creation of a network of trained scientists with appropriate multi-disciplinary skills. The network will provide young researchers with the opportunity to acquire a unique set of skills through international training, and the opportunity to consolidate these skills in an industrial context. Scientific excellence will be supported in a wider context through the establishment of an educational network. Specific measures to be introduced include the establishment of a series of summer schools and workshops, a seminar series, the introduction of modules into undergraduate and post-graduate training programmes, and subject to demand, the provision of industrial training courses.

Employment prospects for network alumni

This network would, over and above the theoretical developments expected, produce trained scientists with a background, which despite a widespread and pressing need, is simply unavailable at present. These would be expected to take leading roles in academia, industry and government. The relevance of the research network to the needs of European industry is confirmed by the strong support expressed for the network’s objectives by industrial organisations including British Aerospace, INEA, CERPD and SINTEF. These observations support the opinion that the employment prospects of network alumni are excellent, over and above the fact that the skills in each of the component disciplines are highly sought after. It is emphasised, moreover, that the requirement for staff with such skills is most certainly not confined to industry and academia. Intelligent embedded systems already pervade many aspects of society and industry, and this trend is clearly set to continue. In view of this widespread, and growing, prevalence, important roles in government certification and safety-related organisations are also envisaged.

 

9. Justification of appointment of young researchers

Training Resources

A total of 224 person months of direct research training for young researchers will be provided by the proposed project. This accords both with the nature of the research objectives and with the scale of the training resources and infrastructure which will be devoted to the project. No young researcher appointment is for longer than 40 months within the 48-month project to provide flexibility in staff recruitment and also to allow a margin for staff loss/re-recruitment. No node will have more than one young scientist to ensure sufficient training support. The network has a healthy mix of pre- and post-doctoral researchers, with Ph.D. students in GU, SU, NUIM and DTU, and post-doctoral researchers in NUST and IJS. The young scientists in this network will have a unique chance to work intensively with more senior staff, as well as with the young scientists at other nodes, because of the high level of commitment in the network towards the common goals.

GU

GU will employ a pre-doctoral researcher for 40 months to work on probabilistic modelling for dynamic systems and the interaction with design methods. This is sufficient time for the researcher to acquire the necessary skills from the network members, help with scientific administration of the network, and to complete a Ph.D. through leading-edge research in probabilistic modelling for engineering systems. The student will be supervised by Dr. Murray-Smith who has extensive experience of supervision of both students and professional research engineers, in both academic and industrial research environments. The PhD. student will benefit from the large research student base in the Computing Science department at GU, from the quality of neighbouring related departments (Statistics, Mech. Eng. and Electronic & Electrical Eng.), and of course from the proximity of SU, the other network node in Glasgow.

SU

SU will also employ a pre-doctoral researcher for 40 months to study modelling and analysis methods for multi-agent systems and investigate their application to renewable energy systems. The minimum period for completion of doctoral studies in the U.K. is 36 months. However, in practice 40 months is a more realistic period within which to pursue research leading to the award of a PhD. The Electronic and Electrical Engineering department at SU has an extensive and well-established doctoral training programme with a sound infrastructure to support training of the appointed young researcher. The young researcher will be formally supervised by Dr. Leithead, who has extensive experience of supervising doctoral students, with additional support provided by Dr. Leith. Moreover, it is envisaged that the students from GU and SU will work closely (as will their supervisors) both in terms of research content and in terms of providing a supportive environment and a strong hub for the network.

NUST

NUST will employ a post-doctoral researcher for 32 months, which provides sufficient time to interact within the network. A post-doc candidate is most ideally suited to the work to be performed at NUST primarily because the suggested tasks are very challenging, and would provide training opportunities at a post-doc level. Moreover, the tasks are closely linked to ongoing work and a post-doc candidate with the right background is expected to integrate more quickly with the existing group than a PhD student. A further benefit is that the existing group at NUST consists mainly of PhD students so a post-doc would be useful resource to the group, and the post-doc candidate would personally benefit from the opportunity to develop research leadership skills working. The post-doc will be supervised by Profs. Tor A. Johansen and Bjarne A. Foss, who have long experience in supervising students and research leadership, both academic and industrial.

NUIM

Despite much recent attention on the study of hybrid dynamical systems, many profound questions relating to their stability remain unanswered. This is due to a number of factors; the challenging nature of these problems, and the fact that there are relatively few researchers in engineering disciplines with the necessary skills to study these problems. In view of this, and with the particular aim of building research competence in this area, recruitment of a pre-doctoral researcher is most appropriate. This satisfies the need to train young researchers in the above area, and best matches the academic interests of the researchers involved in the Irish node of the network. A training period of 40 months within a 48-month network has been selected as it best accommodates the objectives of academic research and the completion of a PhD training programme, while allowing time for staff recruitment and/or staff loss and re-recruitment. The Department of Computer Science, NUIM, has extensive experience of research training in a European Context. The department has been involved in several research networks, and currently hosts a number of non Irish-nationals who are pursuing post-graduate study. Dr. Shorten has supervised students working in the area of Dynamic Systems. In addition, Dr. Shorten is a former Marie Curie fellow and is a co-ordinator of the Marie Curie Fellowship Association in Ireland. He is familiar with past and current problems associated with research training in a host country, and has been involved on a national level in resolving issues of concern to fellows in Ireland. This experience will be invaluable in the initial stages of the project during the initial training and relocation, and in ensuring the successful completion of the project.

DTU

DTU will employ a pre-doctoral researcher for 36 months, the standard period for a Danish Ph.D. The student will acquire skills in Bayesian modelling, and modern, computational implementation of Bayesian inference. Original research will be pursued in the interface between Bayesian statistics and dynamic systesm modelling and control. DTU offers an excellent, modular set of PhD-level courses which will provide a strong background in mathematical modelling. The Ph.D. student will be able to join a very active young research group with a healthy blend of post- and pre-doctoral researchers. DTU will also accept students from elsewhere in the network on short-term visits for specific skill-development in implementation of Bayesian inference, and participation in intensive (week-long) specialist course modules.

IJS, Ljubljana

IJS will employ a post-doctoral researcher for 36 months to experimentally investigate and support the theoretical developments from the other nodes. In view of the need for an initial induction period for the young researchers at other network nodes, and to allow time for the initial development of analysis and design methods, the IJS appointment will commence around 6 months after the appointments at other nodes. The 36 month duration of the post and the appointment at post-doctoral level reflects the considerable experience and skills required to successfully pursue challenging applications work. The researcher will benefit personally from the experience gained in the experimental application of leading-edge ideas and the opportunity to gain significant international research leadership and management experience within the stimulating environment provided by IJS and the training network. The department is an environment where one has opportunity to gain from the practical experience and theoretical background of IJS staff as well as their experience in bridging the gap between theory and practice. The researcher will be able to pursue tasks in a semi-industrial process laboratory containing plant which consists of the following sub-processes: natural gas combustion, gas-liquid separator, heat exchanger, selective catalytic reduction of nitrogen oxides and chemical reactor for neutralisation.

Recruitment Strategy

Vacancies will be well publicised internationally via co-ordinated measures including

It is anticipated that the generous funding being offered, the challenging nature of the work, and the quality of the researchers in the network, are factors which will help to overcome recruitment difficulties associated with a buoyant employment market, and attract scientists with highly sought-after skills.

Promotion of Equal Opportunities

Each of the partners has a strong policy of promoting equal opportunities which is co-ordinated by a dedicated equal opportunities officer/committee. This is reflected in active participation in initiatives such as the U.K. Women Into Science and Engineering (WISE) programme and GU’s Women into Computing scheme. SU has approximately 40% female participation in the control systems research group – high by standards in the subject. The rights of all applicants for employment, irrespective of gender, race, creed, or sexual orientation will be protected. Established guidelines regarding non-gender specific wording of adverts will be adopted and appointment procedures will be open and adhere to equal opportunity guidelines.

Young researchers to be financed by the contract

Participant

Pre-docs (months)

Post-docs (months)

Total

Scientific Specialities

1. GU

40

0

40

M-13,M-01,M-12,M-10

2. SU

40

0

40

M-13,M-01,I-04,I-07

3. NUIM

40

0

40

M-13

4. NUST

0

32

32

M-13, I-04, I-07

5. DTU

36

0

36

M-01,M-13

6. IJS

0

36

36

M-13,I-07

Totals

156

68

224

 

 

10. Training programme

Training Objectives

The primary objective is to provide a core of European scientists with state-of-the-art skills in theoretical and practical methods for the analysis and design of multi-agent control systems. This will be augmented by the wider dissemination of knowledge to European industry and students over and above the network participants.

International Dimension & Special Training Measures

In view of the multi-disciplinary and transnational nature of the subject area of the network, a high proportion of training will be carried out on a network-wide basis. Initially, the training will be more formal and course-orientated but as the research progresses, workshops, secondments, personal discussions and tuition will play an increasingly important training role. Specific network-wide training measures will include:

  1. Network Induction. Soon after the inauguration of the network, an induction workshop will be organised to provide young researchers with an introductory overview of the network’s research topic.
  2. Taught Modules. Local nodes in the network will provide specialist training for the young researchers. Groups of young researchers will be encouraged to travel to appropriate nodes in order to take intensive courses in core subjects together.
  3. Summer schools. Two summer schools will be organised to broaden the skills of the young researchers. This facility will also be available to the wider academic community with international experts invited to give lectures on topics broadly relevant to the networks research area.
  4. International Conferences and Affiliated Workshops. It is expected that each young researcher will actively participate in and give presentations to relevant national and international conferences during the lifetime of the network. In addition, it is intended to organise a number of specialist tutorial and research workshops affiliated to appropriate international conferences.
  5. Academic and Industrial Secondments. To facilitate training and knowledge transfer between research groups, young researchers will be expected to spend significant time at more than one network node. Industrial training will also be offered through secondments to IJS/INEA, and possibly to external collaborators such as BAe, or DaimlerChrysler.
  6. Internet-Based Learning. The use of internet-based methods for distance learning of specialist topics will be explored. Seminars, summer school materials and workshop tutorials will be made available on the web. Young researchers will be encouraged to publish results in the form of tutorials for other members of the network. A "journal club" and discussion group will be established to support joint literature reading, learning, team-work and to generally encourage an active day to day exchange of ideas and results.
  7. Teaching Experience. Opportunities to gain teaching experience will be provided through guest lectures as part of undergraduate and graduate modules, and through the presentation of course modules to industry.
  8. Wider Skills. Each of the network nodes will provide assistance with language skills. Young researchers will be provided with training in presentational, paper-writing and project management skills and will be given responsibility to organise their own network activities. Young researcher meetings and workshops will be an integral part of the network.

Training Environment

Each university involved in the network is a major educational establishment with a well-developed training infrastructure. In addition to the general opportunities to develop communication and management skills during the running of the network and via attendance and presentation at conferences, workshops etc.. Formal courses are also available to help young researchers develop their presentation, management, communication and language skills (substantial support is available in the form of intensive summer courses). English is regularly used as a working language in all nodes. Where appropriate, participation in such programmes would be strongly encouraged. Opportunities to gain teaching experience will be provided through guest lectures as part of appropriate advanced undergraduate degree courses. At a more general level, the organising participants in the network have extensive personal experience of working abroad which will assist in anticipating and understanding the needs of young researchers in a broad sense over and above research and training.

11. Multidisciplinarity in the training programme

The research and training objectives of the network are outlined in Part 1b and Part 10. The achievement of these objectives inherently requires the combination of skills from a number of disciplines, and this is reflected in both the structure of the network, and in the training measures introduced by the network. The research and training network includes two Computing Science departments (GU, NUIM), one Electronic & Electrical Engineering department (SU), a Mathematical Modelling department (DTU), an Engineering Cybernetics department (NUST) and an industrial research node (IJS). This automatically brings a variety of research cultures to the training programme, as well as the essential complementary know-how provided by the different groups. Knowledge transfer between disciplines is central to the ethos of the network, and in addition to day-to-working-level interaction, competence from each node in the team will be combined and incorporated into the training program through extended academic and industrial secondments, workshops and summer schools. This will be augmented by inviting experts from a variety of relevant disciplines to participate in summer schools, which will greatly enhance the training benefit of the network. These benefits will be available directly to network young researchers and members (academic staff), and also to non-network members through summer schools and the provision of courses to universities and industry. This range of multidisciplinary expertise is vital for the success of the proposed project, and the integration of staff from computer science, statistics, control and systems engineering for the analysis and design of multi-agent control systems, is a primary objective of this project. Currently there are very few scientists or engineers who have been exposed to this range of knowledge, and scientific literature. We expect "graduates" of our network to go on and build further bridges between disciplines in future.

12. Connections with industry in the training programme

The network training programme will make extensive use of the strong industrial links which exist to the aerospace, automotive, process, power generation and distribution industries. In particular, the industrial organisations affiliated to the network (British Aerospace, INEA, CERPD, SINTEF) will be utilised to assist in the definition and undertaking of benchmark applications including, where appropriate, the provision of data and assessment of research outcomes. Engineers from industrial partners will be invited to participate in the workshops and summer schools organised by the network. With regard to the training of the young researchers these measures will offer a number of significant benefits including:

  1. Provision of training in the context of industrially relevant problems.
  2. Building an awareness of industrial needs and practice.
  3. Providing exposure to economic drivers of engineering design (at least as important as technical drivers in many cases).

It is emphasised that the industrial input will be of a substantial nature and include secondments to industrial companies (INEA and possibly other collaborators such as British Aerospace and DaimlerChrysler), provision of real data and, if appropriate, access to a prototype commercial wind turbine for field testing. The industrial partners in the network will play a key role in ensuring the delivery of a rounded training programme.

 

13. Financial Information

The financial costs have been carefully estimated, and detailed budgets are available immediately on request.

Personnel costs (Category A)

Each participant has selected a position on the standard wage-scales which is both attractive in the wider European context, and still meets national and institutional constraints on permissible payment levels for given levels of qualification. Added to this, we allowed relocation costs and funding for the scientists to return to their home scientific community every 9 months on average. Where necessary (GU and NUIM) we included Ph.D. fees.

Networking Costs (Category B)

Each participant has been allocated sufficient funds to allow travel to workshops, meetings and conferences. This assumes typically 2 trips per year to meetings/workshops involving an average of 3 researchers, and 2 trips to conferences per year. We have also planned two summer schools, one in Ireland and one in Slovenia. Here we allocated funds to NUIM and IJS to pay for accommodation and catering for all participants from the network to NUIM (4-day school) and IJS (6-day school). Money was also set aside to pay the travel and accommodation costs of inviting 5 external experts to each of the summer schools. External students will cover their own costs.

Further travel money and accommodation was allocated for short exchanges between networks, to improve communication and for specific training purposes. We envisaged an average of two exchanges of up to 3 months, within the network, or to industrial collaborators. These funds would be used to cover extra costs to the young researchers of a further stay abroad.

Consumables, such as standard office supplies, and more importantly, software licenses are also included in the category B costs. This will help cover mission-critical software licenses (typically MATLAB, software development, and Office licenses) which must be provided for the young researchers to pursue their research.

Overheads (Category C)

All participants have included overhead costs at 20% of category A and B.

The percentage of the total dedicated to Category A costs is 721557/1166385 = 61.9% which meets the constraints specified by the Commission. No single country is taking more than 40% of the funding, as stipulated in the guide for proposers.

Financial Information on the network project

Participant

Personnel & mobility costs (A)

(Euro)

Costs linked to networking (B)

(Euro)

Overheads (C)

(Euro)

Totals

(Euro)

1. GU

127342

40800

31775

199917

2. SU

130054

39000

30646

199700

3. NUIM

118585

55187

25900

199672

4. NUST

125301

40964

33253

199518

5. DTU

139448

25300

32830

197578

6. IJS

80827

61380

27793

170648

Totals

721557

262631

182197

1166385

 

References

M. S. Branicky, "Stability of switched and hybrid systems", in Proc., 33rd IEEE Conf. Decision and Control, 1994.

M.S.Branicky, V.S.Borkar, S.K.Mitter, A Unified Framework for Hybrid Control: Model and Optimal Control Theory. IEEE Transactions on Automatic Control, 43, 31-45, 1998.

L.O.Chua, M. Komuro, T.Matsumoto, The double scroll family, parts I & II. IEEE Transactions on Circuits & Systems, 33, 1073-1118, 1986

J. P. Hespania, Logic-Based Switching Algorithms in Control, Ph.D. thesis, Department of Electrical Engineering, Yale University, 1998.

T.A. Johansen, B.A.Foss, Constructing NARMAX models using ARMAX models. International Journal of Control, Vol. 58, pp1125-1153, 1993.

M. I. Jordan, Learning in Graphical Models, NATO Science Series, Kluwer, 1998

Jus Kocijan, John O'Reilly, William E. Leithead, "An integrated undergraduate teaching laboratory approach to multivariable control, IEEE Trans. ed., vol. 40, no. 4, pp. 266-272, 1997.

Jus Kocijan, John O'Reilly, William E. Leithead, "A process control case study in a new multivariable framework", Electrotechnical Review, vol. 64, no. 5, pp. 237-244, 1997.

D.J. Leith, W.E. Leithead, "Strongly Input Constrained Nonlinear Control of a Horizontal Axis Wind Turbine". European Control Conference, Rome, 1995

D.J. Leith, W.E. Leithead, "Performance Enhancement of Wind Turbine Power Regulation by Switched Linear Control", International Journal of Control, Vol. 65, pp555-572, 1996.

D.J. Leith, W.E. Leithead, "Gain-Scheduled & Nonlinear Systems: Dynamic Analysis By Velocity-Based Linearisation Families". International Journal of Control, Vol. 70, pp289-317, 1998.

D.J. Leith, W.E. Leithead, "Survey of Gain-Scheduling Analysis & Design", International Journal of Control, in press, 1999a.

D.J. Leith, W.E. Leithead, Analytic Framework for Blended Multiple Model Systems Using Linear Local Models, International Journal of Control, Vol. 72, pp605-619, 1999b.

D. Liberzon, J. Hespanha,, S. Morse, "Stability of Switched Linear Systems: A Lie Algebraic Condition," tech. rep., Laboratory for Control Science and Engineering, Yale University, 1998.

Y. Mori, Y. Mori, T., Kuroe, Y. "A Solution to the Common Lyapunov Function Problem for Continuous Time Systems," in Proceedings of 36th Conference on Decision and Control, 1997.

S. Morse (1997), (Ed.), Control using Logic-Based Switching, Springer-Verlag

R. Murray-Smith, T. A. Johansen, eds., Multiple Model Approaches to Modelling and Control, Taylor & Francis, 1997.

R. Murray-Smith, T. A. Johansen and R. Shorten, On transient dynamics, off-equilibrium behaviour and identification in blended multiple model structures, European Control Conference, Karlsruhe, 1999.

R. Murray-Smith, Modelling human control behaviour with a Markov-chain switched bank of control laws, IFAC Symposium on Man-Machine Systems, Kyoto 1998.

Rantzer (1999), Dynamic Programming via Convex Optimization, Preprints IFAC World Congress, Beijing

Rantzer and M. Johansson, (1999), Piecewise Linear Quadratic Optimal Control, IEEE Trans. Automatic Control, Vol. 44

R. Shorten and R. Murray-Smith, Side effects of Normalising Radial Basis Function networks, International Journal of Neural Systems, Vol. 7, No. 2, May 1996.

R. Shorten, K.S. Narendra, "Necessary and Sufficient Conditions for the Existence of a Common

Quadratic Lyapunov Function for Two Stable Second Order Linear Systems." Submitted to IEEE Transactions on Automatic Control.

J. Slotine and W. Li, Applied Nonlinear Control, Prentice-Hall, 1991

JM. West, J. Harrison, Bayesian Forecasting and Dynamic Models, Springer-Verlag, 1997

Related policy documents available on the internet.