D. A. Handelman, S. H. Lane, and J. J. Gelfand. Goal-directed encoding of task knowledge for robotic skill acquisition. In Proceedings of the 1991 IEEE International Symposium on Intelligent Control, pages 388--393, Arlington, Virginia, USA, 13--15 August 1991.  Home/Search   Document Not in Database   Summary   Related Articles   Check

....the training information. Methods for supervised learning are primarily devised to store the situation action training pairs in a manner that enables good generalization. Several researchers have used supervised learning techniques to train controllers to perform various tasks (see, for example, [81, 28]) These researchers have used a preexisting expert as a source for the training data. The (usually human) expert supplies a sufficiently large set of situationaction pairs and the network is trained to produce the corresponding action in each situation. This technique is very easy to implement ....

....a human expert to train a controller to solve a pole balancing problem. Widrow draws the parallel between this method and knowledge acquisition from an expert when building an expert system. Once trained, the controller can match the expert s performance or even improve upon it. Handelman et al. [28], for example, show how a connectionist controller can use crude training information from a hand crafted rule based controller to learn a generalized control function that has improved performance. However, such generalized control functions may not always lead to improved performance, especially ....

D. A. Handelman, S. H. Lane, and J. J. Gelfand. Goal-directed encoding of task knowledge for robotic skill acquisition. In Proceedings of the 1991 IEEE International Symposium on Intelligent Control, pages 388--393, Arlington, Virginia, USA, 13--15 August 1991.

....fail. This has motivated the development of intelligent control, which provides different methods involving artificial intelligence features like planification (geometric approach) Can87, CL90, FT87, Lat91, LTJ90] fuzzy and expert systems (rule based approach with structured information) HLG91, OMO 91, vdRvNLD90] neural nets (coding the knowledge into a black box with learning abilities) Kos92, Lee91, Mel90, NSA90, NW90a, NW90b, Wil93] Of course, conventional solutions have a definite advantage: there are theorems which state clearly when the solution exists and when it does ....

D.A. Handelman, S.H. Lane, and J.J. Gelfand. Goal-directed encoding of task knowledge for robotic skill acquisition. In IEEE International Symposium of Intelligent Control, Arlington, VA, 1991.

....will be decreased or left untouched. The preceding discussion is entirely informal and does not pretend to have found the behavior model of any human; it is only based on various experiments (cutting metal plaques with a laser robot [FKB 85] driving a car [Fou90, Luz92a] landing a plane [HLG91] The formal tool resulting from that discussion is rule based incremental control [Luz91] An incremental control law relates u k 1 to u k by: u k 1 = u k ffl k Delta where Delta is a non null positive real and ffl k is an integer in f Gammam; Delta Delta Delta ; mg where m ....

D.A. Handelman, S.H. Lane, and J.J. Gelfand. Goal-directed encoding of task knowledge for robotic skill acquisition. In IEEE International Symposium of Intelligent Control, Arlington, VA, 1991.

....Robot motion has to be considered in its whole and cooperation between the different existing methods could yield an efficient way to solve it. Already some small scale cooperation is available, like mixing neural nets and fuzzy rule based control [Kos92] rule based reasoning and neural nets [HLG91] or geometrical motion planning and non linear control. We have discussed incremental rule based control too, where a theoretical approach proving the existence of a solution copes well with a learning program. These are only first steps and with the rising use of robots, further research is ....

D.A. Handelman, S.H. Lane, and J.J. Gelfand. Goal-directed encoding of task knowledge for robotic skill acquisition. In IEEE International Symposium of Intelligent Control, Arlington, VA, 1991.

....methods are required, however, to ensure that the control function, FW , exhibits interpolation and extrapolation properties that imply good generalization of control. Several researchers have used supervised learning methods to train controllers in supervised learning tasks (see, for example, [150, 53]) These researchers have used a preexisting expert as a source for the training data. The (usually human) expert supplies a sufficiently large set of training pairs specifying the desired actions for various controller inputs, and the controller is trained to produce the corresponding action for ....

....expert to train a controller to solve a pole balancing problem. Widrow draws the parallel between this approach and knowledge acquisition from an expert when building an expert system. Once trained, the controller can match the expert s performance or even 18 improve upon it. Handelman et al. [53], for example, show how a connectionist controller can use crude training information from a hand crafted rule based controller to learn a generalized control function that has improved performance. However, such generalized control functions may not always lead to improved performance, especially ....

Handelman, D. A., Lane, S. H., and Gelfand, J. J. Goal-directed encoding of task knowledge for robotic skill acquisition. In Proceedings of the 1991 IEEE International Symposium on Intelligent Control, pages 388--393, Arlington, Virginia, USA, 13--15 August 1991.

....benefits of fuzzy controllers (see [White and Sofge, 1992] for an introduction) is that it may be simpler to construct control laws using common sense rules than with other methods. Handelman proposes a system that converts rule based fuzzy controllers into appropriate gains for PID controllers [Handelman et al. 1991]. Gelfand consider several tasks that are easy to perform with the use of visual feedback, but assume it is preferable to do the tasks using only proprioceptive feedback [Gelfand et al. 1992] His system is run under visual feedback while a second controller learns the correct mapping using ....

....the advice gets encoded into the control system and you reap the performance benefits. One can imagine a system designer creating an improved controller with the same kinds of suggestions, or rules of thumb. Converting high level rules into controllers is often approached with fuzzy logic ([Handelman et al. 1991] for example) Whether these systems suffer the same problems as local closed loop learners usually depends on the complexity of the rules given by the system designer. Ridding a control system of an incorrect behavior is an interesting problem that has been encountered by other learning systems ....

D. Handelman, S. Lane, and J. Gelfand, "Goal-Directed Encoding of Task Knowledge for Robotic Skill Acquisition," In IEEE International Symposium on Intelligent Control, 1991.

Online articles have much greater impact   More about CiteSeer.IST   Add search form to your site   Submit documents   Feedback  

CiteSeer.IST - Copyright Penn State and NEC