Learning and Transfer of Human Real-Time Control Strategies
Michael C. Nechyba and Yangsheng Xu
The Robotics Institute Carnegie Mellon University, Pittsburgh, PA 15213
In this paper, we address the problem of how to model human real-time control strategy and how to transfer that model to robots or humans. This class of problems is significant to a number of research areas, such as the Intelligent Vehicle Highway System, human-machine interfacing, space telerobotics, and virtual reality. Human models can benefit not only the development of more intelligent control strategies for robots and machines, but can also improve the transfer of human intelligence and skill from expert to apprentice. In this paper, we illustrate a system we developed for modeling human control strategy through the use of flexible cascade neural networks, which adjust the size of the network as part of the training process, and which can be extended with variable activation functions and node-decoupled extended Kalman filtering to achieve faster learning and better error convergence. We implement the method in modeling human real-time driving strategy and show that the HCS models converge to stable behavior, while preserving the differences between individuals’ varying control strategies. We discuss the use of HCS models for transferring skill from human expert to human apprentice; rather than learn directly from a human expert, a HCS model serves as a virtual teacher to a learning apprentice. Finally, we outline on-going research issues and future work related to human control strategy modeling and transfer, including stochastic model validation, and HCS model input selection.