Exploring Social Intelligence through Interactive Multi-Agent Games

As digital technologies become pervasive in day-to-day life, there is an incentive to develop systems that work alongside humans to support their needs. Recent advances in Artificial Intelligence (AI) and Machine Learning (ML) have led to the development systems which can match and even surpass human performance on certain well-defined tasks. However, generalising to unfamiliar tasks, and learning from interactions with humans and other AI agents to solve novel problems, remain open challenge.

Recent work at the intersection of Affective Computing and Multi-Agent Systems has highlighted a lack of "social intelligence" as a critical shortcoming of current state-of-the-art AI systems.

Humans have an awareness of psycho-social factors which underpin our decision-making, such as character and affective state (transient emotions and longer-term mood states). This awareness plays an important role in our unique ability to cooperate, collaborate and learn from one another to solve complex problems. For example, individuals may learn to form cooperative relationships with others based on how trustworthy they perceive them to be. Similarly, affective state is a crucial factor in human decision-making, which influences how we perceive, evaluate and reason about our environment and others in it at any given moment. Our social intelligence allows us to infer affective information from social cues, and reason about the implications of our affective state or the affective state of others.

Socially intelligent AI agents may leverage knowledge about how these human factors influence behaviour to improve their ability to reason about other agents and humans in their environment. However, these factors are often ambiguous, heterogeneous in their presentation, and may vary over time or with the current environment and social context, making them difficult to model.
This research project will explore behavioural cues from human interaction (with other humans or autonomous agents) to better understand the underlying principles of human decision-making, and to inform the development of novel models with complex social behaviour.
The aims of this work are to develop and evaluate techniques for modelling behaviour through interactive agent-based simulation, and to gather and analyse data from empirical study of human-human and human-agent interaction,. To address these aims, the proposed methodology will be centred around the use of interactive multi-agent game environments as research testbeds, designed to support experiments on different mixed-group populations (comprising artificial agents, humans, or a combination of both) within a consistent task setting. To best capture the social dynamics of real-world tasks, this work will focus on "mixed-motive" games, e.g., "Sequential Social Dilemmas", which offer opportunities for players to choose cooperative or adversarial actions throughout.

This project will involve modelling and theoretical analysis of human and agent decision-making strategies. The project will investigate various AI, ML and probabilistic modelling methods to explore the relationships between psycho-social characteristics and observed behaviour in the game environment.