Algorithms - Policy Iterations

Qi := Qp (s,a)

pi(s) := argmaxaeA {Qi-1 (s,a)}

Convergence:

1. Number of iterations less than Value Iterations.

2. Never considers the same policy twice, |A||S|.

RECENT: number of iterations is O(|A||S|/ |S|) [MS, UAI99].

Previous slide

Next slide

Back to first slide

View graphic version