WP2: Models for integrative AI

Task Lead: FBK, PI: Luciano Serafini, Co-PI: Marco Pistore

Development of well-founded approaches for the integration of heterogeneous AI models (including neural, probabilistic, logical, dynamic and decision models) into a single agent; study of the dynamic of these models due to the interaction of the agent with the environment and the intra-agent integration that happens at the social level. 

Effort 30p/m + cascading funds.

Description of work 

The work in this WP aims at providing the foundation of integrative AI, by providing the theoretical context to support the tight integration of heterogeneous AI models, and by defining  computationally workable algorithms for learning, inference and decision making exploiting these integrated models. The work is organized in three strictly connected tasks, which aim respectively at defining integrative models and algorithms (i) for individual AI agents, (ii) for interactive AI agents that act in a work populated by other similar AI agents, and (iii) for evolving societies of artificial and human agents. The models and algorithms developed in this WP will be validated on representative use cases, defined in collaboration with the other WPs of the Spoke.

Task 2.2.1: Integration of heterogeneous AI models. - Task Leader: Luciano Serafini

Design of theoretically well founded models that tightly integrate different AI models. This includes architectures for neuro-symbolic integration, such as the ones already proposed, e.g., LTN and KENN, or the integration of logic and probability in the area of statistical relational learning. Further research direction includes the integration of discrete structures representable by logical theories and continuous structures representable by multivariate continuous distributions. For these integrated models we want to study algorithms that learn the models (parameters) from data and perform inference to support decision making.

Task 2.2.2: Architecture of agents that perceive, learn, plan, reason, and act in an open world -  Task Leaders: Tommaso Campari, Leonardo Lamanna

Development of computationally viable architectures and algorithms for autonomous agents that are able to plan and execute actions that allow them to acquire, and revise their knowledge of the environment where they operate. This environment is supposed to be open and populated by other similar agents, and only partially known by the agent. The acquired knowledge should be represented in suitable integrated AI models (as those developed in task 2.2.1). The key tasks addressed in this task are the following: (i) learn abstract symbolic (discrete) abstract models from high dimensional continuous observations (e.g., images);(ii) plan and execute actions that allow the agent to acquire supervisions for learning models (e.g., execute actions with known effects, or explicitly ask supervision to other agents); (iii) learn how abstract  symbolic actions can be compiled down in sequences of simple operations executable by the agent actuators.

Task 2.2.3: Modeling of a dynamic society  - Task Leader: Marco Pistore

The goal of this task is to study how a society of artificial agents (or mixed artificial and human agents) evolves, taking into account in particular the role of social rules in this evolution. In particular we are interested in the following aspects: (i) if an agent is operating in a society of agents that assumes a set of social rules, how such an agent becomes aware of such a rule and decides to adopt it; (ii) how social rules emerge from the behavior of the single interacting agents; (iii) how global behavior can be induced by imposing/proposing social rules in the society; and (iv) how (the existence of) social rules can be effectively represented as integrated AI models (as those defined in task 2.2.1) in order to enable computationally workable  learning, inference, and decision making algorithms. In all the scenarios we assume that the agents can decide to follow a social rule or they can deliberately decide not to follow them.