Optimistic Planning for Sparsely Stochastic Systems. Lucian Busoniu , Remi Munos, Bart De Schutter, Robert Babuska

(2-pager from ICAPS MCTS workshop)

1. Goal is to plan in settings where there is a restricted number of possible next states in any transition
2. Alg is called Optimistic Planning for Sparsely Stochastic Systems
3. Alg has access to list of all possible next states
4. Bounds are kept for each possible transition, and states with the maximum upper bound are expanded first
5. Analysis is w.r.t. simple regret
6. Bounds propagate up the tree
7. Seems to do rollouts following upper bounds, feels similar to FS^3
8. In their experiments, has better regret, and plans deeper than OLOP (in a discrete version of swingup).  OLOP needs more samples to plan effectively, and so has performance similar to uniform planning.  Computational times are similar
9. By following that expansion rule, alg also maximizes information gained relevant to regret

Advertisements