Success in ICCM

We are proud to announce that these papers were accepted at the International Conference on Cognitive Modeling. The first one will be presented as a talk, and the second as a poster. The conference is in Toronto in July.

Fast Online Reinforcement Learning with Biologically-Based State Representations
- M Bartlett, TC Stewart, J Orchard
This study explored whether the use of biologically inspired grid cells for representing spatial information in a spatial navigation reinforcement learning task. Several Actor-Critic networks were developed, each using a different method for representing the agent’s state (position in space). Parameter optimization identified optimal parameter sets for each network and revealed that whilst most methods performed well once optimized, the network using grid cells did not necessarily require optimization in order to produce optimal performance.

Biologically-Plausible Memory of Continuous-Time Reinforcement Learning
- M Bartlett, N Dumont, M Furlong, TC Stewart
This paper presents a novel Temporal Difference learning rule which operates in continuous time - TD(theta). The rule is implemented in an Actor-Critic network using Legendre Delay Networks for storing historical information, and performance is compared with a network using the discrete-time TD(n) learning rule. The results demonstrate that the novel TD(theta) produces similar behaviour to TD(n), and the theoretical implications and avenues for future research are discussed.