Can Evolutionary Spiking and Non-spiking Neural Networks Yield More Robust Solutions for Multiagent Learning Problems?
Duration: 24 months (2010-2012)
Principal Investigator: Chris Christodoulou
Main Funding Source:University of Cyprus
The research area of reinforcement learning (RL) aims to build autonomous agents that are capable of learning effective behaviours for challenging sequential decision making tasks.
In this work, we explore various aspects of the problem, such as
(i) how to increase the performance of agents in single- and multi-agent environments through hierarchical decomposition of the task into reusable subtasks,
and (ii) how to accelerate learning in game theoretic situations by transforming the payoffs into rewards that motivate the agents to achieve their goal.
In particular, we investigate the evolution of non-spiking neural network controllers in the context of single-agent and multi-agent toy problems. We additionally compare the behaviour of non-spiking agents with spiking ones in a game theoretic scenario known as the Iterated Prisoner’s Dilemma and show that it can be very similar. The case study of the Cyprus problem is investigated in the context of multiagent learning, where we model the various phases and show how a compromise could be achieved in the current phase.
This work concludes with the evolution of learning, where local rules are evolved in single- and multi-agent scenarios. The University of Cyprus was the principal investigator of this project and the University of Birmingham, UK was a partner.