Learning General Policies for Planning through GPT Models

Abstract

Transformer-based architectures, such as T5, BERT and GPT, have demonstrated revolutionary capabilities in Natural Language Processing. Several studies showed that deep learning models using these architectures not only possess remarkable linguistic knowledge, but they also exhibit forms of factual knowledge, common sense, and even programming skills. However, the scientific community still debates about their reasoning capabilities, which have been recently tested in the context of automated AI planning; the literature presents mixed results, and the prevailing view is that current transformer-based models may not be adequate for planning. In this paper, we address this challenge differently. We introduce a GPT-based model customised for planning (PLANGPT) to learn a general policy for classical planning by training the model from scratch with a dataset of solved planning instances. Once PLANGPT has been trained for a domain, it can be used to generate a solution plan for an input problem instance in that domain. Our training procedure exploits automated planning knowledge to enhance the performance of the trained model. We build and evaluate our GPT model with several planning domains, and we compare its performance w.r.t. other recent deep learning techniques for generalised planning, demonstrating the effectiveness of the proposed approach.