Date of Award
Program or Major
Doctor of Philosophy
W Thomas Miller, III
Redundant (i.e., under-determined) systems can not be trained effectively using direct inverse modeling with supervised learning, for reasons well out-lined by Michael Jordan at MIT. There is a "loop-hole", however, in Jordan's preconditions, which seems to allow just such an architecture. A robot path planner implementing a cerebellar inspired "habituation" paradigm with such an architecture will be introduced. The system, called ARTFORMS, for "Adaptive Redundant Trajectory Formation System" uses on-line training of multiple CMACS. CMACs are locally generalizing networks, and have an a priori deterministic geometric input space mapping. These properties together with on-line learning and rapid convergence satisfy the loop-hole conditions. Issues of stability/plasticity, presentation order and generalization, computational complexity, and subsumptive fusion of multiple networks are discussed.
Two implementations are described. The first is shown not to be "goal directed" enough for ultimate success. The second, which is highly successful, is made more goal directed by the addition of secondary training, which reduces the dimensionality of the problem by using a set of constraint equations. Running open loop with respect to posture (the system metric which reduces dimensionality) is seen to be the root cause of the first system's failure, not the use of the direct inverse method. In fact, several nice properties of direct inverse modeling contribute to the system's convergence speed, robustness and compliance.
The central problem used to demonstrate this method is the control of trajectory formation for a planar kinematic chain with a variable number of joints.
Finally, this method is extended to implement adaptive obstacle avoidance.
Rudolph, Franklin J., "A neural network-based trajectory planner for redundant systems using direct inverse modeling" (1992). Doctoral Dissertations. 1716.