In recent years, various studies have been carried out to assess whether Large Language Models (LLMs) possess different reasoning capabilities, including those required in automated planning. Typically, these studies provide the LLM with a planning domain and a problem, specified by an initial state and a goal, and require the LLM model to generate a plan solving the problem. Despite this common configuration, such studies significantly differ in the used models, the information provided to the model, the possible involvement of symbolic planners, and the experimental approaches used for the evaluation. Motivated by the growing interest in LLMs and in the understanding of their reasoning abilities, in this work we offer a concise review of recent studies on using LLMs for planning. We outline the main research trends and discuss their most notable findings. Furthermore, we identify key challenges and highlight critical aspects to consider when evaluating a LLM in terms of learning to plan and generating solution plans.

Transformer-based architectures have revolutionized natural language processing and represent a significant promise for advancing policy learning in generalized planning tasks. In particular, PlanGPT demonstrated remarkable results in generating solution plans in different planning domains. For each domain, it is trained on a domain-specific dataset composed of randomly generated planning problems and corresponding solution plans. This work proposes a novel extension to the PlanGPT framework by incorporating macro-actions, high-level actions that encapsulate sequences of primitive actions, which are a well-established technique in classical planning known to improve planning efficiency by guiding exploration through shortcuts. Leveraging this concept, we investigate the impact of macro-actions on the learning process of PlanGPT and whether they mitigate the occurrence of violated preconditions caused by complex object relationships, particularly in domains where these interactions frequently lead to invalid plans due to precondition violations. Experimental results indicate that integrating macroactions improves coverage in several challenging domains and reduces generation time, highlighting the enhanced learning capabilities of PlanGPT when supported by macro-actions.

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.

Most ML-based applications for COVID-19 assess the general conditions of a patient trained and tested on cohorts of patients collected over a short period of time and are capable of providing an alarm a few days in advance, helping clinicians in emergency situations, monitor hospitalised patients and identify potentially critical situations at an early stage. However, the pandemic continues to evolve due to new variants, treatments, and vaccines; considering datasets over short periods could not capture this aspect. In addition, these applications often avoid dealing with the uncertainty associated with the prediction provided by machine learning models, potentially causing costly mistakes. In this work, we present a system based on Recurrent Neural Networks (RNN) for the daily estimate of the prognosis of COVID-19 patients that is built and tested using data collected over a long period of time. Our system achieves high predictive performance and uses an algorithm to effectively determine and discard those patients for whom RNN cannot predict the prognosis with sufficient confidence.

During 2020 and 2021, managing limited healthcare resources and hospital beds has been a fundamental aspect of the fight against the COVID-19 pandemic. Predicting in advance the length of stay, and in particular identifying whether a patient is going to stay in the hospital longer or less than a week, can provide important support in handling resources allocation. However, there have been significant changes in terms of containment measures, virus diffusion, new treatments, vaccines, and new variants of SARS-CoV-2 during the last period. These changes pose several conceptual drift issues that can limit the usefulness of machine learning in this context. In this work, we present a machine learning system trained and tested using data from more than 6000 hospitalised patients in northern Italy, distributed over almost two years of pandemic. We show how machine learning can be effective even by analysing data over this long period of time, also exploiting a model that predicts the patient's outcome in terms of discharge or death. Furthermore, learning from data that also consider deceased patients is a common issue in predicting the length of stay because they have severe conditions similar to patients with a long stay period, but may actually have a very short duration of hospitalisation. For this purpose, we present a method for handling data from alive and deceased patients, exploiting more patient records, increasing the robustness of the model and its performance in this task. Finally, we investigate the features that are most relevant to the prediction of the simplified length of stay.