Automatic Learning of Action Priorities

S.-A. Akramifar and G.-R. Ghassem-Sani (Iran)


AI planning, machine learning, artificial intelligence, agents 1-


Traditional AI planners often suffer from their poor efficiency. There are many choice points in the planning process, but lack of information precludes proper decision. In this paper, we introduce a new method for adding automatic learning capability to a forward planning system. Our idea is based on a dynamic voting algorithm to choose the best action to proceed to the next state. In every planning cycle, applicable actions (i.e. those actions whose preconditions are satisfied in the current world state) vote to, and compete with each other. As a result of this voting, gradually more useful actions are chosen. This idea has been applied to the blocks world domain, and the preliminary results show significant improvement in the performance of the planner, particularly for larger problems.

Important Links:

Go Back