Choosing goals as internal actions in hierarchical AI planning

The author of “A note on goal-based hierarchical RL” says it combines an agent-centric general value function construction with their work on hierarchical hidden Markov models from 25 years earlier. The caveat: no experiments yet. A reply points to “Probabilistic Hierarchical Goal Network Planning with UCT,” saying it shares the assumption that goals can be chosen as an agent’s internal actions. That planning paper describes breaking high-level goals into ordered subgoals. In its experiments on larger problems, the authors report that their compressed search converged more quickly and outperformed their asymptotically optimal search.