Learning general policies for planning through GPT models

Transformer-based architectures, such as BERT, GPT and T5, have achieved remarkable results across various Natural Language Processing (NLP) tasks. Beyond these linguistic capabilities, these Large Language Models (LLMs) exhibit varying degrees of factual knowledge, common sense reasoning, and even programming capabilities. However, their effectiveness in performing logical inference and automated planning remains an open question. Recent attempts to apply LLMs to Classical Planning have produced mixed results. In this thesis, we tackle this challenge by introducing PlanGPT, a GPT-based model trained from scratch on solved planning instances to learn a general policy for Classical Planning. By leveraging domain-specific training data and incorporating automated planning knowledge, PlanGPT can generate solution plans for unseen problems within the same domain, demonstrating good coverage and performance relative to other deep learning approaches. However, there are no formal guarantees of validity and PlanGPT can produce invalid plans that fail to meet all goals or contain actions with unsatisfied preconditions. To mitigate these problems, we propose two approaches. First, we incorporate a validator directly into the generation process, which allows us to prune invalid partial plans on the fly and generate valid solutions. Second, we combine PlanGPT with a plan-repair planner, LPG, which refines invalid or incomplete candidate plans into fully valid solutions. Our empirical evaluations across diverse Classical Planning domains confirm the efficacy of these strategies. Ultimately, this work demonstrates the potential of integrating learned policies with model-based reasoning.