Skip to main content


Algorithms to control Markov decision processes

Trainee Achievements

Algorithms to control Markov decision processes

Trainee Weinstein has developed general planning algorithms to control complex, high-dimensional Markov decision processes. The project focuses on effective sample-based planning in the face of challenges including: high-dimensionality, drift, discrete system changes and stochasticity, all hallmark challenges for important problems, such as humanoid locomotion. To ensure broad applicability, domain expertise is assumed to be minimal. In order to make the method responsive, computational costs must scale linearly with the number of samples. The model is a receding-horizon open-loop planner that employs cross-entropy optimization for policy construction. Simulations demonstrate near-optimal decisions in a small domain and effective simulated locomotion in several challenging humanoid control tasks. (Open-Loop Planning in Large-Scale Stochastic Domains, to appear in the Proceedings of the Association for the Advancement of Artificial Intelligence.)