Machine Learning Control For Highly Reconfigurable High-Order Systems.

摘要

This Grant addressed four focus areas toward advancing the state-of-the-art in learning control theory of robust and adaptive non-equilibrium control of highly nonlinear, higher-order, reconfigurable systems: 1. Extend Approximate Dynamic Programming (ADP) techniques to control of nonlinear, multiple time scale, non-affine systems in an Adaptive Control framework.; 2. Develop solution techniques for Markov Decision Problems (MDP) that scale to continuous state and control spaces with constraints; 3. Extend MDP techniques to solve multi-agent co-ordination and control problems in a decentralized fashion. 4. Develop solution techniques that scale to continuous state-space Partially Observable Markov Decision Problems (POMDP) and their multi-agent generalizations. The work produced the first significant results in the nonlinear control of multiple time-scale control in the last 25 years, and additionally made significant contributions to the analysis and control of systems that are non-affine in control, and non-minimum phase. The work also developed sampling based feedback planning techniques for the solution of Markov Decision Problems (MDP) and Partially Observed MDPs (POMDP).