论文信息 - Goal directed model inversion

Goal directed model inversion

A new neural network technique for model inversion called goal directed model inversion (GDM) is presented. It allows the system to produce an inverse model in a goal directed manner. The major advantage of an inverse model created in this matter is that it can adapt to unexpected changes in the system with which it must interact. As an example of the GDMI technique, a simple kinematic controller was built for a simulated robotic arm with three degrees of freedom. The system was trained by presenting a sequence of goals of increasing difficulty in some required region of space. As the controller was trained, its ability to extrapolate correct control actions to new distant goals increased.<<ETX>>

Silvano Colombano | Michael Compton | M. Bualat

[1] Richard S. Sutton,et al. Neuronlike adaptive elements that can solve difficult learning control problems , 1983, IEEE Transactions on Systems, Man, and Cybernetics.

[2] Michael I. Jordan,et al. Forward Models: Supervised Learning with a Distal Teacher , 1992, Cogn. Sci..

[3] Charles C. Jorgensen,et al. Distributed memory approaches for robotic neural controllers , 1990 .

[4] Hamid R. Berenji,et al. Refinement of Approximate Reasoning-based Controllers by Reinforcement Learning , 1991, ML.

[5] Michael I. Jordan,et al. Generic constraints on underspecified target trajectories , 1989, International 1989 Joint Conference on Neural Networks.

[6] B. Widrow,et al. The truck backer-upper: an example of self-learning in neural networks , 1989, International 1989 Joint Conference on Neural Networks.