Approximate Dynamic Programming Using Random Sampling

点击翻译按钮获取中文摘要

AbstractProposal Number: ECCS-0824077Proposal Title: Approximate Dynamic Programming Using Random Sampling PI Name: Atkeson, Christopher G. PI Institution: Carnegie-Mellon UniversityThe objective of this research is to develop approximate dynamicprogramming methods for the control of high performance nonlinearsystems. The approach is to use spatially local models of keyfunctions such as a function that represents future costs over anentire task or mission (the value function), and a function that tellsthe system what to do in each situation (the policy). Humanoid robotswill be used to evaluate the nonlinear control design techniques andcompare them with other approximate dynamic programming approaches.Intellectual MeritKey contributions of this work will be: 1) Developing an integratedapproach for dynamic planning based on approximate dynamic programmingfor high performance systems. 2) Showing how to combine many fastgreedy local planners to produce globally optimal solutions. 3)Showing how to represent knowledge along explicit trajectories, whichincreases the feasible sparseness of this approximate dynamicprogramming method. 4) Showing how to create an adaptiverepresentation using random sampling of states. 5) Showing how to uselocal models, and compare them to global parametric functionapproximators. 6) Showing how to use fast trajectory optimization tospeed up approximate dynamic programming.Broader impactsA specific impact of this work will be to enable high performance(near optimal) control in areas like robotics, transportation, andenergy. A more general societal payoff is industrial processes,machines, and robots that are easier to program, perform better andwaste fewer resources.