Dr David Mguni

Profile

David Mguni is a Lecturer in Computer Science at Queen Mary University of London. His research focuses on how intelligent systems can make decisions, coordinate with others, and improve their learning under uncertainty. His work brings together reinforcement learning, game theory, meta-learning, stochastic control and statistical learning, with a particular interest in multi-agent and adaptive AI systems.

A central theme of his research is understanding how learning systems can adapt to new problems, make effective decisions with limited or uncertain information, and interact with other decision-makers. His work spans both theoretical and empirical research, with applications including autonomous and multi-agent systems, mathematical reasoning, scientific modelling and decision support. 

David holds a PhD in Applied Mathematics from University College London and previously studied at the University of Cambridge and the University of London. Before joining Queen Mary, he was a Principal Scientist at Huawei, where he led the Multi-Agent Systems research team in London.

Teaching

Distributed Systems ECS656U / ECS796P 

Distributed systems underpin much of modern computing, from cloud services and distributed databases to peer-to-peer networks and large-scale online applications. The module examines how collections of networked computers can coordinate to behave as a coherent system despite delays, failures and the absence of a single global view. 

The module develops the fundamental ideas needed to reason about distributed computation, including communication between distributed processes; time, ordering and synchronisation; logical and vector clocks; consistency; distributed snapshots; replication and fault tolerance; consensus; and decentralised data structures. Students encounter important algorithms and protocols including Lamport clocks, Chandy-Lamport snapshots, Paxos, Raft and distributed hash tables such as Chord. 

A central aim of the module is to connect formal concepts to concrete system-design questions: what can different machines know about one another, what happens when messages are delayed or machines fail, when agreement is possible, and how systems can continue to operate despite partial failures.

Students develop both an understanding of the core algorithms and the ability to reason systematically about the behaviour, limitations and design trade-offs of distributed systems.

Risk and Decision-Making for Data Science and AI - ECS7005P

This module examines how decisions should be made when information is uncertain, incomplete or potentially misleading.

Students study probability and Bayesian reasoning, quantitative risk, causal reasoning, Bayesian networks, decision analysis and expected utility. The module considers how evidence should change beliefs, how different forms of risk should be communicated and compared, and how decisions can be made when possible outcomes carry different costs and benefits. 

A central theme is the distinction between prediction, explanation and decision-making. Students examine why correlations or highly accurate predictive models do not automatically provide the information required for good decisions, and why causal assumptions, prior information, uncertainty and stakeholder preferences can matter. 

The module combines conceptual reasoning with practical modelling. Students construct and analyse Bayesian networks and decision models using tools such as PyAgrum, investigating how evidence propagates through probabilistic models and how beliefs can be combined with utilities to support decisions. 

Applications are drawn from areas including healthcare, engineering, finance and artificial intelligence. The broader aim is to equip students to critically evaluate claims about risk and use probabilistic and decision-theoretic tools to reason effectively under uncertainty.

Research

Research Interests:

David's research is broadly concerned with learning and decision-making under uncertainty, particularly in settings where intelligent systems must learn, interact and adapt over time.

One strand of his work focuses on reinforcement learning, game theory and multi-agent learning. This includes understanding how multiple learning agents cooperate, compete and coordinate; how useful behaviours and equilibria emerge through learning; and how decision-making can remain efficient, robust and safe when agents have different objectives, information or capabilities. 

A second strand focuses on meta-learning and adaptive or self-improving AI. This work studies how learning systems can use experience across tasks to adapt more effectively to new problems, how aspects of the learning process itself can be learned, and how systems can autonomously acquire the information and experience needed to improve. This includes problems in few-shot learning, continual adaptation and the generation or selection of informative learning experiences.

 His research also studies the relationship between the structure of learning problems and the behaviour of learning algorithms. This includes understanding how properties of tasks, games and dynamical systems influence learning and optimisation, and how this structure can be exploited to design or select more effective learning methods. 

A further strand concerns probabilistic and stochastic models of complex systems, drawing on stochastic control, Bayesian methods and stochastic differential equations to study learning and decision-making in uncertain and evolving environments. 

These themes increasingly extend to large language models and agentic AI systems, including reasoning, mathematical problem solving, multi-agent interaction and adaptive learning. Across these areas, the broader objective is to develop intelligent systems that can adapt reliably to new tasks, environments and other decision-makers rather than being restricted to a fixed problem or training setting.

See David Mguni’s research profile pages including details of research publications, and live grants.

Supervision

I supervise research projects broadly concerned with machine learning, artificial intelligence and decision-making under uncertainty. I am most interested in supervising projects that pose a clear scientific question about how or why an intelligent system behaves as it does. 

Areas of particular interest include: 

• Reinforcement learning and multi-agent systems: cooperation and competition, equilibrium learning, safe and robust reinforcement learning, reward design, exploration and coordination between heterogeneous agents. 

• Meta-learning and self-improving AI: learning how to learn, adaptive learning, few-shot learning, continual improvement and autonomous generation or selection of useful learning experiences. 

• LLM reasoning and agentic systems: mathematical reasoning and theorem proving, strategy and plan evaluation, uncertainty and abstention, and communication between multiple LLM-based agents.
  
  
• Games, optimisation and learning algorithms: the behavioural and geometric structure of games and optimisation algorithms, algorithm selection and efficient equilibrium computation.
  

Projects can range from focused empirical investigations using existing models and datasets to more technically ambitious work involving new learning algorithms, mathematical analysis or theoretical results. Projects combining theoretical and empirical approaches are also welcomed.