Methodologies
Driving better approaches to collective behaviour research
To address collective behaviour challenges, we combine methodologies from different fields such as computer science, mathematics, physics and more.
Agent-based simulations (ABMs)
ABMs are a powerful simulation-based approach. Combining decision-making rules with simulated sensory input, agents individually decide how to move, change their mental state or alter their environment. Physical properties, such as realistic movement dynamics, can also be taken into account.
ABMs provide a detailed and infinitely flexible way of modelling. Significant experience is needed to engineer agent-based simulations that accurately capture the scenario modelled. They are powerful tools for answering fundamental questions about the behaviour of social groups, and allow us to predict large-scale dynamic behaviours emerging from many individual interactions. For example, to optimise agricultural processes, we have constructed models that simulate the behaviour of millions of pollinating bee-agents in detail as they navigate and forage among virtual flowering plant agents.
Our group uses these models to understand insect societies, human crowds and consumer behaviour, among many other applications.
Individual-based models (IBMs)
Our lab uses IBMs to understand complex behaviour in biological and human societies, from the movement of cells via social insects to human crowds. We also apply them to robot swarms, technical systems in which complex interactions between many individuals play a central role.
IBMs are more restrictive than general agent-based models (ABMs), but their fundamental advantage is that they are simpler, which means they can not only be executed very efficiently but, more importantly, they are amenable to mathematical analysis. IBMs naturally bridge computer simulation and mathematical modelling. Their simplicity allows us to apply methods from statistical physics to analyse their properties and obtain theoretical insights that a simulation alone cannot produce. As a result, we can explore complex scenarios and vast parameter spaces in ways that ABMs cannot.
Casting a real-world system as an individual-based model requires drastic abstractions. The art is to find the salient aspects and strip away the rest.
We address this challenge of abstraction with multi-model approaches. We start from relatively realistic ABMs, reduce these to IBMs and finally capture their core properties in mathematical models. The power of such methods is that they combine the exploratory freedom of simulation models with the certainty of understanding that only a mathematical model can give.
Our team has the knowledge and experience to ensure that multi-models are congruent and can capture relevant aspects of the real-world system at all levels.
Game theory
Game theory allows us to predict how one action will affect the choices made by others.
The central concept in game theory is the equilibrium, which arises when no agent in the system has an incentive to unilaterally change their actions – stabilising group behaviours. This is a very powerful method that allows us to predict how a scenario may play out when agents are rational.
We have a particular interest in evolutionary game theory (EGT), which studies the dynamics of behaviours in groups of agents that can learn. EGT lets us analyse which equilibria are more plausible, enabling predictions in social, natural and artificial systems.
Our lab applies game theory to understand tensions between individual motivations and social outcomes. The fundamental question is: How can a desirable outcome for a group be achieved when the events are driven by independent individuals?
Areas of work here include the division of labour, function of reputation systems and impact of incentives on social endeavours.
Reinforcement learning (RL)
RL is a mathematical framework that describes how individuals learn to act in an unknown environment. This learning is derived from experiencing the results of interactions with this environment. Conceptually, this is very close to how people and animals often learn by trial-and-error rather than by complex reasoning.
RL is an extremely active area of machine learning and the foundation of many groundbreaking AI solutions. However, its origins lie in earlier attempts to understand animal behaviour. We revitalise this research direction.
Being one of the core methods in modern AI, large amounts of research have been devoted to RL and we have achieved a good understanding of its theory. Using RL as a modelling approach (rather than as a machine learning method) affords us access to a rich toolbox of theory for explaining collective behaviour.
When groups of individuals work together, the experiences and the learning of many agents interacts so this framework, which was originally designed for individual agents, is expanded to multi-agent reinforcement learning (MARL).
Our aims are twofold: we use MARL as a tool to explain collective behaviour in real-world domains and at the same time we contribute to advancing the general theory of MARL – an approach path at the cutting-edge of mainstream AI research.
Optimisation models
Optimisation models provide a fundamental alternative to agent-based or individual-based approaches. The central concept is the best or most desirable outcome for a system, which leads to a top-down modelling approach where we start by defining a target outcome and then derive the behaviours that achieve it (or, more commonly, parameterisations for a loosely defined behaviour).
This complements the agent-based bottom-up approach, where we first define the behavioural rules for agents and subsequently simulate the model to investigate the consequences of these behaviours. As a result, we can address questions about the efficiency of complex behaviours that are beyond the reach of agent-based models alone.
AI-based behaviour monitoring
We use AI-based methods, mostly from computer vision and computer listening, to monitor the behaviour of social groups in the real world as the basis of our research.