Predicting novel paths to goals by a simple, biologically inspired neural network

Although recurrent networks can be used as content addressable memories, they can also be used as sequence prediction systems. Because problem solving can often be viewed as a sequence prediction problem, we hypothesize that such networks can be used as problem solvers. There are many aspects to problem solving. Here we concentrate on a single but important aspect, goal finding without search. Using a highly simplified, clearly prototypical version of this problem, a sparsely connected recurrent network successfully predicts novel paths to reach a goal.