Evaluation of an Entropy-based Online Sequence Learning Algorithm in Nonstationary Zero-Sum Games

The Hong Kong University of Science and Technology
Department of Computer Science and Engineering

FYT Presentation and Demonstration

Title: "Evaluation of an Entropy-based Online Sequence Learning Algorithm 
       in Nonstationary Zero-Sum Games"

by

Miss YAN, Jialei

abstract:

Constructing agents that can learn and adapt to changing environments is a
key challenge in intelligent systems, especially in multiagent domains,
with the presence of other intelligent agents which are also learning and
adapting. Recently, an entropy-based online sequence learning algorithm,
called Entropy Learning Pruned Hypothesis (ELPH) space, has been proposed
for fast learning in nonstationary environments. This paper presents our
evaluation of ELPH in zero-sum games that simulate nonstationary
multiagent environments. With a properly selected observation history
length of seven and a pruning threshold of 0.5, the ELPH player learns and
adapts to stochastic and nonstationary deterministic agents quickly. ELPH
plays at Nash equilibrium in self-plays of zero-sum games. However, it
loses by a noticeable amount in plays against simulated aggressive human
players. We will analyze the reasons behind the wins and losses.


Date		:	28 July 2008, Monday

Time		:	2:30pm to 3:30pm

Venue		:	Room 3501

Advisor		:	Prof. D.Y. Yeung

2nd Reader	:	Dr. S.C. Cheung