Extending GOLOG with an effective UCT-like MCTS planning capability

Determine how to extend the GOLOG action-language framework so that it provides an effective and practical planning method comparable to UPOM, specifically by incorporating UCT-like Monte Carlo Tree Search over operational models with simulations of nondeterministic outcomes to guide decisions at nondeterministic choice points.

Background

The paper contrasts its RAE+UPOM framework—where planning operates directly on operational models using UCT-like Monte Carlo Tree Search and simulated rollouts—with logic-based approaches such as GOLOG. While GOLOG includes control constructs and can in principle use forward search at nondeterministic choices, the authors argue that existing GOLOG-based systems lack an effective, practical planning method like UPOM that leverages MCTS over simulated outcomes.

The authors explicitly state uncertainty about how GOLOG could be extended to achieve this capability, framing an unresolved question about whether and how GOLOG can be endowed with an efficient, simulation-driven UCT-like planning mechanism analogous to UPOM.

References

However, no work based on GOLOG provides an effective and practical planning method such as UPOM, which is based on UCT-like Monte Carlo Tree Search and simulations over different scenarios. It is not clear to us how GOLOG could be extended in such a direction.

— Deliberative Acting, Online Planning and Learning with Hierarchical Operational Models (2010.01909 - Patra et al., 2020) in Section 2 (Related Work), Logic Based Approaches