Foreword
Over the past decade, Artificial Intelligence (AI) has made remarkable breakthroughs, particularly in the realm of deep learning and foundation models - sub-symbolic machine learning approaches that leverages deep neural networks with hundreds of billions parameters. These models are often called black boxes as their human interpretation and understanding is very limited. This technology has been instrumental in enhancing interaction, perception, and natural language processing, sometimes even surpassing human capabilities. As a result, some researchers have begun to equate AI with deep learning and foundation models. However, I believe this is a significant misconception.
AI encompasses far more than just sub-symbolic machine learning; it includes symbolic (i.e., human-understandable) modeling, search algorithms, and reasoning techniques - all vital aspects of human intelligence that extend beyond machine learning, and can potentially utilize it to enhance algorithm performance and model accuracy.
Planning and acting are intrinsic human abilities. Even young children naturally plan and act, learning from the consequences of their actions in an environment and refining their skills as they grow. Machines have not yet reached human-level proficiency in planning and acting, as well as in their integration with learning, leaving considerable room for advancement and improvements in autonomous intelligent systems.
This book serves as a crucial milestone in the study of planning, acting, and learning, exploring how these intelligent features can be effectively combined and integrated to improve the performance of intelligent systems. The authors, Malik Ghallab, Dana Nau, and Paolo Traverso, are three outstanding scientists and researchers who have achieved significant recognition and visibility within the AI international scientific community. This is the third book they have written on the subject: the first focused on planning, while the second explored the interaction between acting and planning. This third book marks an important step forward by also addressing the intersection of acting, planning, and learning. It discusses Deterministic State-Transitions, Hierarchical Task Networks, Probabilistic, Non-deterministic, Hierarchical-Refinement, and Temporal Models, while also considering Robotic Motion and Manipulation. Additionally, it explores the emerging capabilities of Large Language Models and how they can be applied in this field, a very recent and relevant topic at the intersection between sub-symbolic and symbolic AI.
The book is not only a valuable reference for scientists working in the area but also serves as a textbook for graduate students, offering a clear, comprehensive, and well-organized catalogue of techniques and algorithms for domain modeling, plan construction, and execution, as well as the integration of learning in all these activities. I have no doubt that I will recommend it in my courses and use it as a personal reference.
Michela Milano
University of Bologna