We introduce a joint framework for model leaning and planning in stochastic domains with continuous state and action spaces with non-Gaussian transition noise. It is efficient in large spaces with large amounts of data because (1) local models are estimated only when the planner requires them; (2) the planner focuses on the most relevant states to the current planning problem; and (3) the planner focuses on the most informative and/or high-value actions. Our theoretical analysis shows that the expected difference between the optimal value function of the original problem and the value of the policy we compute vanishes sub-linearly in the number of actions we test, under mild assumptions. We show empirically that multi-modal transition models are necessary if the underlying dynamics is not single-mode, and our algorithm is able to complete both learning and planning within minutes for a stochastic pushing problem in simulation given more than a million data points, as a result of focused planning.
from cs.AI updates on arXiv.org http://ift.tt/2atId3l
via IFTTT
No comments:
Post a Comment