Solving multi-stage games with hierarchical learning automata that bootstrap

Hierarchical learning automata are shown to be an excellent tool for solving multi-stage games. However, most updating schemes used by hierarchical automata expect the multi-stage game to reach an absorbing state at which point the automata are updated in a Monte Carlo way. As such, the approach is infeasible for large multi-stage games (and even for problems with an infinite horizon) and the convergence process is slow. In this paper we propose an algorithm where the rewards don't have to travel all the way up to the top of the hierarchy and in which there is no need for explicit end-stages.