A partially recurrent gating network approach to learning action selection by reinforcement

We describe a neural network approach to the problem of reactive navigation, using a simulated mobile robot. Specifically, it is shown that complementary reinforcement backpropagation learning can be a means for modular networks to acquire different navigation related skills concurrently, Further, it is demonstrated that a partially recurrent net can function as a gating network to coordinate the reinforcement learning across modules and across time steps. In effect, the recurrent gating network performs action selection by choosing developing experts to make control decisions in the context of previous actions in the temporally extended domain.