Information Theory for Distributed AI (INFORMED-AI)

Artificial intelligence (AI) is on the verge of widespread deployment in ways that will impact our everyday lives. It might do so in the form of self-driving cars or of navigation systems optimising routes on the basis of real-time traffic information. It might do so through smart homes, in which usage of high-power devices is timed intelligently based on real- time forecasts of renewable generation. It might do so by automatically coordinating emergency vehicles in the event of a major incident, natural or man-made, or by coordinating swarms of small robots collectively engaged in some task, such as search-and-rescue.

Much of the research on AI to date has focused on optimising the performance of a single agent carrying out a single well-specified task. There has been little work so far on emergent properties of systems in which large numbers of such agents are deployed, and the resulting interactions. Such interactions could end up disturbing the environments for which the agents have been optimised. For instance, if a large number of self-driving cars simultaneously choose the same route based on real-time information, it could overload roads on that route. If a large number of smart homes simultaneously switch devices on in response to an increase in wind energy generation, it could destabilise the power grid. If a large number of stock-trading algorithmic agents respond similarly to new information, it could destabilise financial markets. Thus, the emergent effects of interactions between autonomous agents inevitably modify their operating environment, raising significant concerns about the predictability and robustness of critical infrastructure networks. At the same time, they offer the prospect of optimising distributed AI systems to take advantage of cooperation, information sharing, and collective learning.

The key future challenge is therefore to design distributed systems of interacting AIs that can exploit synergies in collective behaviour, while being resilient to unwanted emergent effects. Biological evolution has addressed many such challenges, with social insects such as ants and bees being an example of highly complex and well-adapted responses emerging at the colony level from the actions of very simple individual agents!

The goal of this project is to develop the mathematical foundations for understanding and exploiting the emergent features of complex systems composed of relatively simple agents. While there has already been considerable research on such problems, the novelty of this project is in the use of information theory to study fundamental mathematical limits on learning and optimisation in such systems. Information theory is a branch of mathematics that is ideally suited to address such questions. Insights from this study will be used to inform the development of new algorithms for artificial agents operating in environments composed of large numbers of interacting agents.

The project will bring together mathematicians working in information theory, network science and complex systems with engineers and computer scientists working on machine learning, AI and robotics. The aim goal is to translate theoretical insights into algorithms that are deployed onreal world applications real systems; lessons learned from deploying and testing the algorithms in interacting systems will be used to refine models and algorithms in a virtuous circle.