Generalized maze navigation: SRN critics solve what feedforward or Hebbian nets cannot

Previous papers have explained why model-based adaptive critic designs-unlike other designs used in neurocontrol-have the potential to replicate some of the key, basic aspects of intelligence as seen in the brain. However, these designs are modular designs, containing "simple" supervised learning systems as modules. The intelligence of the overall system depends on the function approximation abilities of these modules. For the generalized maze navigation problem, no feedforward networks-MLP, RBF, CMAC, etc. or networks based on Hebbian learning have good enough approximation abilities. In this problem, one learns to input a maze description, and output a policy or value function, without having to relearn the policy when one encounters a new maze. This paper reports how we solved a very simple but challenging instance of this problem, using a new form of simultaneous recurrent network (SRN) based on a cellular structure which has some interesting similarity to the hippocampus.