Background
Training a reinforcement learning agent from scratch for each new edge environment is expensive — in time, data, and compute. Yet edge networks vary enormously: a placement policy tuned for a dense urban deployment may perform poorly in a suburban network with different server topologies, user mobility patterns, and service mixes. Transfer Reinforcement Learning (TRL) addresses this directly, by asking whether knowledge acquired in one domain can be reused — or fine-tuned at low cost — in another.
This project frames cross-network generalisation as the core research problem. Can a placement agent trained on one city’s edge network be deployed, with minimal retraining, in another city? Can a policy learned in simulation transfer to a real deployment? These questions are central to making edge AI commercially viable at scale.
Research Challenge
Reinforcement Learning (RL) has shown significant promise for optimising service placement and resource management in edge computing environments. However, most existing approaches require extensive retraining whenever the deployment environment changes. This creates a major barrier to deploying intelligent edge management systems across different cities, network infrastructures and application domains.
This project investigates how Transfer Reinforcement Learning (TRL) can enable service placement agents to generalise across diverse edge environments. Rather than learning from scratch for every new deployment, agents will be trained to reuse knowledge acquired in one network and adapt rapidly to new topologies, mobility patterns, resource constraints and service demands.
A key challenge is understanding which aspects of a learned placement strategy are transferable and which are specific to a particular environment. The research will explore challenges such as zero-shot and few-shot adaptation, domain shift, policy transfer, simulation-to-reality transfer, knowledge representation and robustness to previously unseen operating conditions. Particular attention will be given to identifying the factors that influence successful transfer and determining how transferable knowledge can be exploited to reduce training costs while maintaining high-quality placement decisions.
Example application domains include transferring service placement policies between different smart cities, adapting AI-driven transportation systems to new urban environments, deploying edge intelligence in rapidly changing disaster-response scenarios, and enabling AI service management platforms to scale across multiple network operators. The work will involve designing and evaluating transfer learning strategies across a diverse set of simulated edge environments and digital twin models.
Topics
Transfer Reinforcement Learning (TRL); Zero-Shot and Few-Shot Learning; Cross-Domain Generalisation; Reinforcement Learning; Edge Computing and Edge AI; Digital Twins; Simulation-to-Reality Transfer; Adaptive AI Systems; Meta-Learning and Policy Transfer; Autonomous Service Placement
Impact
One of the defining goals of modern artificial intelligence is to create systems that can transfer knowledge between tasks and environments rather than repeatedly learning from scratch. Achieving this capability is essential for building scalable and economically viable AI systems that can be deployed across diverse real-world settings.
This project addresses a fundamental challenge in adaptive AI: how to enable intelligent decision-making systems to generalise across different environments while minimising retraining effort. The outcomes could contribute to more robust and deployable AI solutions for smart cities, telecommunications networks, transportation systems and future autonomous infrastructures. More broadly, the project explores a key step towards AI systems that can learn efficiently, adapt rapidly and operate effectively in previously unseen environments.