Use this URL to cite or link to this record in EThOS: http://ethos.bl.uk/OrderDetails.do?uin=uk.bl.ethos.423157
Title: Fuzzy and tile coding approximation techniques for coevolution in reinforcement learning
Author: Tokarchuk, Laurissa Nadia
ISNI:       0000 0001 3534 4346
Awarding Body: Queen Mary, University of London
Current Institution: Queen Mary, University of London
Date of Award: 2005
Availability of Full Text:
Access through EThOS:
Access through Institution:
Abstract:
This thesis investigates reinforcement learning algorithms suitable for learning in large state space problems and coevolution. In order to learn in large state spaces, the state space must be collapsed to a computationally feasible size and then generalised about. This thesis presents two new implementations of the classic temporal difference (TD) reinforcement learning algorithm Sarsa that utilise fuzzy logic principles for approximation, FQ Sarsa and Fuzzy Sarsa. The effectiveness of these two fuzzy reinforcement learning algorithms is investigated in the context of an agent marketplace. It presents a practical investigation into the design of fuzzy membership functions and tile coding schemas. A critical analysis of the fuzzy algorithms to a related technique in function approximation, a coarse coding approach called tile coding is given in the context of three different simulation environments; the mountain-car problem, a predator/prey gridworld and an agent marketplace. A further comparison between Fuzzy Sarsa and tile coding in the context of the nonstationary environments of the agent marketplace and predator/prey gridworld is presented. This thesis shows that the Fuzzy Sarsa algorithm achieves a significant reduction of state space over traditional Sarsa, without loss of the finer detail that the FQ Sarsa algorithm experiences. It also shows that Fuzzy Sarsa and gradient descent Sarsa(λ) with tile coding learn similar levels of distinction against a stationary strategy. Finally, this thesis demonstrates that Fuzzy Sarsa performs better in a competitive multiagent domain than the tile coding solution.
Supervisor: Not available Sponsor: Not available
Qualification Name: Thesis (Ph.D.) Qualification Level: Doctoral
EThOS ID: uk.bl.ethos.423157  DOI: Not available
Keywords: Electronic Engineering
Share: