Efficient Parameter Optimisation using Genetic Programming

Dr Joao C. F. Pujol (Leverhulme Visiting Fellow)
Work in collaboration with Riccardo Poli

Parameter optimisation problems present themselves in a variety of important domains ranging from engineering to artificial intelligence, from mathematics to the biological sciences, in areas such as function optimisation, system identification, control, machine learning, design, and many others. Traditionally, in order to solve such problems using computers, one has to specify a representation for the parameters of the system, and then an algorithm that will perform the optimisation of these parameters.

In this seminar, a new idea to parameter optimisation called Parameter Mapping Approach (PMA) will be discussed. PMA's distinctive feature is that, instead of using the standard array-based representation, it encodes the parameters to be optimised as functions that are evolved using Genetic Programming (GP). The appeal of the novel representation is that these functions may be interpreted as learning algorithms tailored by evolution to the problem being addressed. Also, the novel representation is, in principle, independent of the size of the problem being tackled.

In the talk results will be presented to show, that although PMA is the first technique ever to enable GP to do classical function optimisation, its performance is at least as good, if not superior, to well known specialised algorithms.

Friday, 1st October 2004
Seminar in Room 4B.531 at 3pm

Co-evolution versus Self-play TD Learning for Acquiring Position Evaluation in Go

Thomas Runarsson and Simon Lucas

Two learning methods for acquiring position evaluation for small Go boards are studied and compared. In each case the function to be learned is a position-weighted piece counter and only the learning method differs. The methods we study are Temporal Difference Learning (TDL) using the self-play gradient-descent method and Evolutionary Learning (EL), in particular co-evolution using an Evolution Strategy (ES).

The two approaches are compared with the hope of gaining a greater insight into the problem of searching for “optimal” zero-sum game strategies. The main result is that TDL consistently outperforms EL, converging to more powerful players, and in a fewer number of games played.

The seminar will include a short tutorial overview of how TDL (a type of reinforcement learning) can be applied to learn game strategies.

Friday, 23rd September 2004
Seminar in Room 4B.531 at 3pm

[top of page]


  Monday, 27 September 2004