posted on 2023-05-26, 09:42authored byOllington, R, Vamplew, P
This paper presents a new algorithm for goal-independent Q-learning. The model was tested on a simulation of the Morris watermaze task. The new model learns faster than conventional Q-learning and experiences no interference when the goal location is moved. Once the new location has been discovered the system is able to navigate directly to the platform on subsequent trials. The model was also tested on watermaze tasks involving barriers. The presence of barriers did not affect the acquisition of "one-trial" learning. While presented as a navigational and mapping technique, the model could be applied to any reinforcement learning task with a variable reward structure.
History
Department/School
School of Computing
Publication status
Published
Event title
The 2nd International Conference on Computational Intelligence, Robotics and Autonomous Systems