Game-theoretic logic learning in scientific domains

Add to your list(s) Download to your calendar using vCal

• Stephen Muggleton - Imperial College
• Tuesday 27 July 2010, 14:00-15:00
• Small public lecture room, Microsoft Research Ltd, 7 J J Thomson Avenue (Off Madingley Road), Cambridge.

If you have a question about this talk, please contact Microsoft Research Cambridge Talks Admins.

This talk has been canceled/deleted

Abstract: This talk is motivated by an attempt to model the process of automated scientific discovery in terms of the theory of competitive and collaborative games. For instance, at the object-level within Systems Biology interactions between a host and pathogen can be modelled as a form of adaptive competition. Conversely at the meta-level groups of experimental scientists can be viewed as conducting a collaborative game involving the proposal and refutation of hypotheses by experimentation. Some relevant concepts from Game Theory are briefly reviewed. We then introduce a formalism called Game-Theoretic Logic Programs (GTLPs), which allow modelling of multi-player strategies based on an adaptation of McCarthy`s situation calculus. Lastly we propose an approach to machine learning such strategies. We argue that the approach, called Pre-emptive Strategy Learning (PSL), represents a departure from traditional forms of Machine Learning. Usually Machine Learning is conceived in terms of extracting patterns from a database of past experiences with the aim of predicting future outcomes. By contrast, experimental choices in new areas of science might be compared to the problem BP faced during the oil spillage in the Caribbean. The absence of relevent historical data dictates a new approach. It is envisaged that PSL will machine learn strategies by sampling projected future events from a GTLP description of actions with associated probabilities and costs. In this way, machine learning could be applied iteratively within cycles of experiment planning. Although the development of GTL Ps and PSL would have direct and immediate effect within automated scientific discovery tasks, we believe that the approach should also have broader application within other areas of computer science in which situations and actions have associated uncertainties and costs.

This talk is part of the Microsoft Research Cambridge, public talks series.

This talk is included in these lists:

This talk is not included in any other list

Note that ex-directory lists are not shown.