Evolution of internal dynamics for neural network nodes

Abstract

Most artificial neural networks have nodes that apply a simple static transfer function, such as a sigmoid or gaussian, to their accumulated inputs. This contrasts with biological neurons, whose transfer functions are dynamic and driven by a rich internal structure. Our artificial neural network approach, which we call state-enhanced neural networks, uses nodes with dynamic transfer functions based on n-dimensional real-valued internal state. This internal state provides the nodes with memory of past inputs and computations. The state update rules, which determine the internal dynamics of a node, are optimized by an evolutionary algorithm to fit a particular task and environment. We demonstrate the effectiveness of the approach in comparison to certain types of recurrent neural networks using a suite of partially observable Markov decision processes as test problems. These problems involve both sequence detection and simulated mice in mazes, and include four advanced benchmarks proposed by other researchers.