On the Computational Complexity of Model Reconciliations

Abstract

A longstanding objective in classical planning is to synthesize policies that generalize across multiple problems from the same domain. In this work, we study generalized policy search-based methods with a focus on the score function used to guide the search over policies. We demonstrate limitations of two score functions --- policy evaluation and plan comparison --- and propose a new approach that overcomes these limitations. The main idea behind our approach, Policy-Guided Planning for Generalized Policy Generalization (PG3), is that a candidate policy should be used to guide planning on training problems as a mechanism for evaluating that candidate. Theoretical results in a simplified setting give conditions under which PG3 is optimal or admissible. We then study a specific instantiation of policy search where planning problems are PDDL-based and policies are lifted decision lists. Empirical results in six domains confirm that PG3 learns generalized policies more efficiently and effectively than several baselines.

On the Computational Complexity of Plan Verification, (Bounded) Plan-Optimality Verification, and Bounded Plan Existence

“…Many vital tasks are centered on plan optimality verification, e.g., the task of model reconciliation, of plan post-optimization, and of domain learning. The first one is to change a planning problem's domain with the least number of changes so as to turn a plan into an optimal solution, which is Σ p 2 -complete (Sreedharan, Bercher, and Kambhampati 2022). The second one is concerned with whether a plan can be further optimized by removing some redundant actions from it, which is NP-complete in both classical planning (Fink and Yang 1992;Nakhost and Müller 2010) and POCL planning (Olz and Bercher 2019).…”

Learning Planning Domains from Non-redundant Fully-Observed Traces: Theoretical Foundations and Complexity Analysis

“…Domain learning with expert knowledge is related to both domain modification (Lin and Bercher 2021) and model reconciliation (Sreedharan, Bercher, and Kambhampati 2022), which has the same setting as domain modification but additionally requires that the given plan is optimal in the modified domain. The constraint that the domain to be found can have at most k modifications from the base domain can also be expressed as a propositional logical formula.…”

Was Fixing This Really That Hard? On the Complexity of Correcting HTN Domains

“…To our best knowledge only a few approaches exist that provide further AI-based support. Lindsay et al (2020), for example, refined an inaccurate hybrid domain to capture the environment more accurately, and Sreedharan et al (2020) revised a dialogue domain via model reconciliation (Sreedharan, Chakraborti, and Kambhampati 2021; Sreedharan, Bercher, and Kambhampati 2022). Lin and Bercher (2021) also studied the complexity of finding corrections to a flawed domain model provided a plan that shall be a solution but currently is not.…”