Learning Subjective "Cognitive Maps" in the Presence of Sensory-Motor Errors

In this paper we present a new version of our previous work on a maze learning animat. Its sensory/motor capabilities have been extended and modified so that they are more biologically plausible than before. The animat's learning architecture is based around a hybrid RBF Neural Network/Evolutionary Strategy implementation of an Adaptive Heuristic Critic. We conduct experiments in which the animat either acquires persistent but undetectable internal errors in its sensory equipment, or operates in an environment where undetectable factors influence motor actions. We also observe the effects of random sensory errors on the usefulness of the information which the animat acquires. Through interactions with its environment the animat learns a subjective “cognitive map” which is a fusion of the features in its surroundings, the path to a goal state, and the errors/environmental influences which it cannot directly detect. We find that despite the subjective nature of the map it remains useful under quite high levels of error/distortion in our experiments.