Towards an integration of learning and reasoning in agent and multi-agent systems

The introduction of artificial intelligence (AI) systems in human daily life represents one of the most significant transformations of our time. This ever-increasing presence requires the development of sophisticated systems capable of tackling and solving more complex tasks. These systems are no longer simply machines, but computational entities (or rather agents) that need to adapt, interact, and understand both their environment and human beings. Several new research directions have emerged to develop these increasingly specialized and capable systems. These include the integration of sub-symbolic techniques with symbolic ones, known as Neuro-Symbolic (NeSy) AI. This integration enables the creation of heterogeneous systems, which we define as systems with a variety of capabilities (such as learning and reasoning) that may cope more successfully with more complicated tasks or better operate in specific circumstances, such as constrained environments. In this work, we illustrate how different ways of integration between symbolic and subsymbolic modules allow the development of hybrid agent approaches. We refer to hybrids as both the interaction between sub-symbolic and symbolic agents that result in what we call hybrid multi-agent systems, and the interaction between sub-symbolic and symbolic modules within a single agent that results in what we call hybrid agent systems. The main objective is to demonstrate, through different integration architectures, how these hybrid systems can be used to obtain systems capable of: (i) integrating and using external knowledge to improve their performance and possibly adapt to changing environments, (ii) exploiting different capabilities in order to achieve greater completeness and complexity, and (iii) at the same time responding to multiple requirements. These requirements include, for instance, the need for less complex systems that are more transparent, ethical, and understandable for humans, as well as the need for systems that are sustainable from a computational and energy point of view, but also systems that are more efficient in terms of the amount of training data requirements. Accordingly, in this thesis, we will focus on developing hybrid agents and hybrid multi-agent systems to highlight the advantages of combining learning and reasoning. In addition, we will provide some metrics to accurately assess the potential benefits of this form of integration.