The Department of Computer Science at the University of Cyprus cordially invites you to the PhD Defense entitled:

Studies in Reinforcement Learning and Adaptive Neural Networks

Speaker: Mr. Vassilis Vassiliades
Affiliation: University of Cyprus, Cyprus
Category: PhD Defense
Location: Room 148, Faculty of Pure and Applied Sciences (FST-01), 1 University Avenue, 2109 Nicosia, Cyprus (directions)
Date: Tuesday, July 14, 2015
Time: 09:30-10:30 EET
Host: Chris Christodoulou (

Intelligent biological organisms are characterised by their ability to learn from past experience and adapt to their unpredictable environments. Creating software or robots that exhibit similar characteristics would not only be beneficial and valuable, but essential for achieving mankind's boldest dreams. This endeavour, however, presents a number of difficult problems and addressing them all is not an easy task. In this work we specifically investigate adaptation in dynamic environments by exploring two subfields of artificial intelligence research, namely reinforcement learning (RL) and neural networks (NNs). In our first study we deal with multiagent RL (MARL) in a game theoretical situation, where individual agents are self-interested, but in order to maximize their returns they must cooperate. We show that by evolving the reward function of simple RL agents, they perform significantly better on this task. The second study deals with how to accelerate learning in structured MARL tasks by investigating synergies between hierarchical RL and MARL algorithms. We introduce two algorithms for hierarchical MARL and demonstrate that they perform significantly better than their non-hierarchical and non-multiagent counterparts in a partially observable multiagent domain. In the third study we consider the problem of creating adaptive agents by evolving their learning rules. We introduce an approach where both the learning rule and the agent controller are represented as NNs. We show that the evolved learning rules have significantly better performance than a well-known RL algorithm in partially observable versions of three stationary tasks and a nonstationary one. The final study introduces a new type of artificial neuron called "switch neuron" capable of gating all but one of its incoming synaptic connections. We additionally introduce a way of making these neurons modulate other switch neurons and present appropriate switch NN architectures capable of achieving optimal adaptation in (i) arbitrary, nonstationary, binary association problems, and (ii) discrete T-maze domains where continuous behavioural exploration is needed.

Short Bio:
Vassilis Vassiliades is a PhD candidate at the Department of Computer Science of the University of Cyprus under the supervision of Associate Professor Chris Christodoulou. He completed his undergraduate studies at the same department in 2007 and his MSc studies in Intelligent Systems Engineering at the School of Computer Science of the University of Birmingham, UK in 2008. His research interests lie in the field of Bio-Inspired Artificial Intelligence, focusing on reinforcement learning, neural networks and neuroevolution.

  Other Presentations Web:
  Colloquia Web: