This project augments the Blocks-World Hierarchical Agent with look-ahead state evaluation. The description of the original agent applies to this one. The main difference is that look-ahead is performed in the middle of the three problem spaces that it uses.
Hierarchical task composition via subgoaling
Internally simulates external environment including an i/o link
agent incorporates both operator subgoaling/means ends analysis with reinforcement learning. All search control knowledge (operator evaluation rules) are removed from blocks-world-operator-subgoaling and instead there are RL rules supplemented with rules to compute reward, both in the top state and the substate. Implemented for four blocks.
This agent is a modified version of the simple blocks-world agent that uses means-ends analysis and operator subgoaling (first used in General Problem Solver (GPS)).
Means-ends analysis involves proposing operators that can achieve part of the goal. Thus, some operators will be proposed even if they do not apply to the current state. If an operator is selected that can not apply, an operator no-change impasse arises. In that substate, the goal is to achieve a state in which the impassed...