Tasks

Task 6.1: Modelling social cognition, collaboration and teamwork (Task Lead: IIIA-CSIC)

To study the modelling of agent’s cognitive capabilities that integrate individual knowledge and behaviour (possibly arising from model-free approaches) with knowledge available to and from other agents (possibly obtained at different times and from different perspectives). This also includes studying the foundations, techniques, algorithms and tools for designing social AI systems. To achieve that, agents should have the capability of understanding others, reason about them (for example have Theory of Mind) and be able to act in a team. We will investigate novel algorithms to perform on-line team formation. Such algorithms will allow to dynamically assemble teams of agents (and possibly humans) to complete tasks that are requested to be serviced along time. We will also consider aspects such as: fostering diversity within teams in terms of cognitive abilities, personality and gender and manage the dynamics of teams. We will also cater for agents’ motivations by considering their preferences about taking part in tasks as well as how they perceive others, namely their potential team-mates. Finally, our algorithms will also cater to the perception and past observations of assessing agents about teams. Therefore, besides being capable of on-line operation, an important novelty of our team formation algorithm will stem from considering the modelling of the perceptions about others of both working agents and assessing agents, namely from considering social cognition aspects.

Task 6.2: Theoretical models for cooperation between agents (Task Lead: UOX)

In this task we will use economic paradigms to study and advance the foundations, techniques, algorithms and tools for collaborative decision making by social agents. As AI agents act on behalf of people, a first crucial issue is to model and elicit their preferences and in particular to aggregate and mediate preferences of multiple stakeholders in a fair manner. The second issue is that self-interested agents often need to be given additional incentives to motivate them to execute their tasks faithfully. While economics has shown many impossibility results, multi-agent systems often allow creating artificial settings that allow more powerful mechanisms. In particular, machine learning allows tailoring mechanisms to the particular preferences of agents, using a technique called automated mechanism design. Another opportunity arises from the fact that most AI optimization algorithms now use randomization which invalidates many impossibility results from economics, allowing for example truthful protocols for social choice and budget-balanced truthful auctions.

Task 6.3: Learning from others (Task Lead: VUB)

We will study the foundations, techniques, algorithms and tools for social learning. We will address the key question of who should learn from whom, and what should be learned. The setting of a single learning agent guided by another agents or by humans has been extensively studied in the past few years. It can exhibit different forms such as learning from demonstrations, advice and imitation learning. Through shaping and interaction, learning can be guided, and when the shaping obeys certain conditions, such a potential function, the learning can still be guaranteed to converge to an optimal behavior. When the learning agent needs to perform a task which is not the same as the ‘teacher’ agent, or when the capabilities of the learning agent or different, one might use transfer learning approaches. In this task we will investigate the research questions arising from placing multiple learning agents in a social context with other agents. How can these agents be efficiently guided in their joint learning process? And how to explore the link between representation, optimization and learning in a multi-agent learning setting. We will also consider federated learning, where agents collaborate to learn a joint model while keeping their individual data private.

Task 6.4: Emergent Behaviour, agent societies and social networks (Task Lead: CNR)

In this task we will look at the society level, studying the foundations, techniques, algorithms and tools for modeling and designing complex social structures, organizations and institutions. AI has already influenced the work in socio-technical systems (STS), in cyber-physical systems (CPS) and in multi-agent systems (MAS), that is, the most successful approaches for understanding, controlling and maintaining systems where large populations of natural and artificial entities interact in multiple ways with rich information exchanges and mutual behavioural dependencies. Some of the different approaches to modelling social systems that we will consider in TAILOR are self-organization, evolutionary game theoretical paradigms and agent-based simulations. We will also consider normative systems that arise from the collaborative agreements of the members of the society on the norms that regulate their interactions.

Task 6.5: Synergies Industry, Challenges, Roadmap on social AI system (Task leader: TNO)

Task 6.6: Fostering the AI scientific community on the theme of social AI (Task leader: IST-UL)

This task aims at:

  1. Promoting activities such as bilateral/multilateral meetings among scientists,

  2. Promoting and supporting student visits,

  3. Organising workshops on Social AI, promoting the area

  4. Organising summer schools

  5. TAILOR conference and other common activities related to the area of Social AI