Exploiting Model Uncertainty Estimates for Safe Dynamic Control Learning

Model learning combined with dynamic programming has been shown to be e ective for learning control of continuous state dynamic systems. The simplest method assumes the learned model is correct and applies dynamic programmingto it, but many approximators provide uncertainty estimates on the t. How can they be exploited? This paper addresses the case where the system must be prevented from having catastrophic failures during learning. We propose a new algorithm adapted from the dual control literature anduse Bayesian locally weighted regression models with stochastic dynamic programming. A common reinforcement learning assumption is that aggressive exploration should be encouraged. This paper addresses the converse case in which the system has to reign in exploration. The algorithm is illustrated on a 4 dimensional simulated control problem.

• Type: Conference
• Year: 1996
• Where: NIPS
• Authors: Jeff G. Schneider
• Keywords: Dynamic Programming, NIPS 1996, NIPS 2007, State Dynamic Systems, Stochastic Dynamic Programming