Stabilizing Human Control Strategies through Reinforcement Learning

M. C. Nechyba and J. A. Bagnell
Abstract
Humans are, and for the foreseeable future remain our best and only example of true intelligence. In comparison, even advanced robots are still embarrassingly stupid. Consequently, one popular approach for imparting intelligent behaviors to robots and other machines abstracts models of human control strategy (HCS), learned directly from human control data. This type of approach can be broadly classified as "learning through observation." A competing approach, which builds up complex behaviors through exploration and optimization over time, is reinforcement learning. We seek to unite these two approaches and show that each approach, in fact, complements the other. Specifically, we propose a new algorithm, rooted in reinforcement learning, for stabilizing learned models of human control strategy. In this paper, we first describe the real-time driving simulator which we have developed for investigating human control strategies. Next, we motivate and describe our framework for modeling human control strategies. We then illustrate how the resulting HCS models can be stabilized through reinforcement learning and finally report some positive experimental results.
M. C. Nechyba and J. A. Bagnell, "Stabilizing Human Control Strategies through Reinforcement Learning," Proc. IEEE Hong Kong Symp. on Robotics and Control, vol. 1, pp. 39-44, 1999 (284 kb).