On Learning Action Costs from Input Plans

2024-08-20

arXiv.org

AI Summary

This paper addresses the problem of unknown action costs in classical planning, formally introducing the novel task of *learning action costs from input plans*: given a set of unlabeled optimal (or k-optimal) plans, infer action costs such that all input plans are optimal (or k-optimal) under the learned cost model. To solve this, we propose LACFIP^k, an algorithm integrating integer linear programming modeling, feasibility-driven iterative optimization, cost-space pruning, and explicit k-optimality constraints. We provide a theoretical proof of its finite-step convergence. Experiments across multiple planning domains demonstrate that LACFIP^k perfectly recovers the correct action cost ranking (100% accuracy), significantly outperforming existing baselines. Our work establishes a new paradigm for inverse planning and enhances model interpretability by enabling cost inference directly from observed optimal behavior.

Technology Category

• Planning, Routing, and Scheduling: Learning for Planning and Scheduling
• Search and Optimization: Learning to Search
• Constraint Satisfaction and Optimization: Constraint Learning and Acquisition

Application Category

• Economics, Online Markets and Human Computation: Cost models of using LLMs in production systems
• Search and Retrieval-Augmented AI: Web learning to rank, online learning, and counterfactual learning for ranking
• Graph Algorithms and Modeling for the Web: Algorithms and analysis for incomplete, noisy, or partially observed Web-related graphs

Abstract

Most of the work on learning action models focus on learning the actions' dynamics from input plans. This allows us to specify the valid plans of a planning task. However, very little work focuses on learning action costs, which in turn allows us to rank the different plans. In this paper we introduce a new problem: that of learning the costs of a set of actions such that a set of input plans are optimal under the resulting planning model. To solve this problem we present $LACFIP^k$, an algorithm to learn action's costs from unlabeled input plans. We provide theoretical and empirical results showing how $LACFIP^k$ can successfully solve this task.

Problem

• Learning action costs from unlabeled input plans
• Ensuring input plans become optimal solutions
• Ranking different plans through cost optimization

Innovation

• Learning action costs from unlabeled plans
• Algorithm LACFIP^k solves cost optimization
• Ensures input plans become optimal solutions