Oleg Sinyavskiy

Method of reinforcement learning of spiking neural network that controls robot or virtual agent is described. Using spiking neurons as key elements of a network allows us to exploit spatial and temporal structure of input sensory information. Teaching of the network is implemented with a use of reinforcement signals that come from the external environment and reflect the success of agent’s recent actions. A maximization of the received reinforcement is done via modulated minimization of neurons’ informational entropy that depends on neurons’ weights. The laws of weights changes were close to modulated synaptic plasticity that was observed in real neurons. Reinforcement learning algorithm was tested on a task of a resource search in a virtual discrete environment.