Risk-sensitive optimal control for stochastic recurrent neural networks

Abstract

As a continuation of our study, this paper extends our research results of optimality-oriented control from deterministic recurrent neural networks to stochastic recurrent neural networks, and presents a new theoretical design for the risk-sensitive optimal control of stochastic recurrent neural networks. The design procedure follows the technique of inverse optimality, and obtains risk-sensitive state feedback controllers that guarantee an achievable meaningful cost for a given risk-sensitivity parameter.