Co-Evolutionary Learning

Prof. Jordan Pollack
Computer Science Department
Brandeis University

Colloquium & AIRG/AIDG Presentation
Friday, December 12, 1997
11 a.m.
Kaven Hall 204

We work on learning agents who face an environment (a teacher, training set or fitness function) which is dynamically modulated as the agent adapts - often as the result of other adapting agents. We use the term "co-evolution" to refer to an inspiring process from nature which supports our goal of growing systems of great complexity. Many scientists have worked on this in different applications, such as prisoners dilemmas, predator/prey, code warriors, virtual robots, self-playing game-learners, with wildly varying success. We will present some of our results - in game learning, pattern classification, and language induction. In particular I focus on a simple and surprising hill-climbing replication of Tesauro's 1992 self-learning backgammon player. Although self- learning often fails through early convergence to unexpectedly stable situations, for backgammon the simplest possible learning apparently succeeded. To understand why, we describe the teacher and learner as agents in a learning game we call the "Meta Game of Learning" (MGL), which suffers from many collusive equilibria we call "Mediocre Stable States." What are the features of backgammon, and of life itself, which prevent MSS's in the MGL?

Maintained by webmaster@wpi.edu
Last modified: Sep 27, 2006, 16:05 EDT