In the week beginning 20th April, we will have three talks from Msci students, abstracts below:
Tuesday 21st April, 2pm, 423 SAWB: Demian Till (Further Investigating Novelty Search)
Thursday 23rd April, 2pm, 422 SAWB Dimitar Petrov (Improving Touchscreen Typing Using Back-of-Device Grip Interactions) and Martin Bevc (Predicting the Outcome of Tennis Matches From Point-by-Point Data)
Further Investigation Novelty Search – Demian Till
By focussing on objectives, optimisation algorithms tend to waste a lot of time exploring areas of the search space around local optima. In 2008, Lehman and Stanley  introduced a new algorithm named ‘novelty search’, which completely ignores the objective and instead directs search based on how much candidates’ behaviours differ from those of previously discovered candidates. In 2011, Lehman and Stanley  showed that novelty search significantly outperformed a standard genetic algorithm in the challenging problem of simulated bipedal locomotion. However, the behavioural characterisation used by Lehman and Stanley  arguably slipped in a standard fitness function through the back door. This paper demonstrates that novelty search still outperforms ‘objectivebased’ search when using a measure of novelty that contains no information about distance travelled. We then introduce a modification of the novelty search algorithm which we show to outperform the original algorithm on the problem of simulated bipedal locomotion. Finally, we investigate the effects of combining novelty search with objectivebased search.
 Lehman, Joel, and Kenneth O Stanley. “Exploiting OpenEndedness to Solve Problems Through the Search for Novelty.” ALIFE5 Aug. 2008: 329336.
 Lehman, Joel, and Kenneth O Stanley. “Abandoning objectives: Evolution through the search for novelty alone.” Evolutionary computation19.2 (2011): 189223.
Improving Touchscreen Typing Using Back-of-Device Grip Interactions – Dimitar Petrov
Abstract: Typing on touchscreen keyboards is inherently inaccurate as users tend to touch locations offset from their intended target. Offsets are user-specific and can further differ for a given user between postures (left-hand, right-hand, two-hand typing). On the other hand, back-of-device interaction has been used to predict screen touches on randomised abstract targets. We propose a new approach where unique offset models are learned for each posture and back-of-device is used to predict posture. Offset models are learned using linear regression while classification is achieved by SVMs and GPs. The device used is a regular smartphone extended with a capacitive sensor on the back.
Predicting the Outcome of Tennis Matches From Point-by-Point Data – Martin Bevc
Tennis is one of the most popular sports in the world and the format of the
game has made it one of the most heavily traded sports in betting markets.
With opportunities for big profits, interest in accurate predictions is high
among professional traders and amateur gamblers.
Traditionally research in predictions of outcomes of tennis matches has focused
on aggregating a lot of historical data and using simple statistical methods to
compute winning probabilities.
The talk will explore modeling tennis matches with Markov chains,
alternative approaches to making outcome predictions
such as simulations, making predictions from in play point by
point data and by sampling points in situations where available data is sparse.
Performance of these methods and techniques used in previous research will