From exploration to planning

Abstract. Learning and behaviour of mobile robots faces limitations. In reinforcement learning, for example, an agent learns a strategy to get to only one specific target point within a state space. However, we can grasp a visually localized object at any point in space or navigate to any position in a room. We present a neural network model in which an agent learns a model of the state space that allows him to get to an arbitrarily chosen goal via a short route. By randomly exploring the state space, the agent learns associations between two adjoining states and the action that links them. Given arbitrary starting and goal positions, route-finding is done in two steps. First, an activation gradient spreads around the goal position along the associative connections. Second, the agent uses state-action associations to determine the actions leading to ascend the gradient toward the goal. All mechanisms are biologically justifiable.