Previous 
Up 
Next
The focus of this project is artificial neural systems (ANS) design with an emphasis on sensory-motor integration and goal-directed operation. Here we look at behaviour from a top-down point of view. We consider an autonomous system, e.g. a mobile robot, with sensory and motor capabilities existing in and interacting with an environment where there is consumables, dangers, possibly in the form of other autonomous agents.
The important thing is that our agent also has drives and "feelings" of generalized pain and pleasure. Methods to "program" such an autonomou agent by supplying it with the appropriate drives are studied. The controlling ANS should be capable of learning a stimulus-response mapping that optimizes the reward-punishment function reasonably well. We further want to develop and study mechanisms to control attention, learning of "appetitive and avoidance behaviours" and making appropriate behaviour selection etc. Orienting responses, working memory and temporal credit assignment are other important problems considered.
An ongoing masters thesis project (Björn Hedin) is currently trying out the ANSim simulator for control of a simulated mobile robot. The robot itself is simulated using the simulator Flakey (by K. Konolige at Stanford Research Insitute). For the moment, the task studied is collision avoidance. Different learning strategies are tested (i) The original Flakey fuzzy controller is used and the ANS controller sits beside to observe and learn the input-output mapping; (ii) Reinforcement learning using only the ANS controller. A report is planned for the spring of 1995.
Next