Varnie Ketheeswaran
Pronouns: He/Him
Academic and research departments
Centre for Vision, Speech and Signal Processing (CVSSP), School of Computer Science and Electronic Engineering.About
My research project
Smart Urban Representation for Visual AI Infrastructure Management (SURVAI)I am a PhD researcher in the Centre for Vision, Speech and Signal Processing (CVSSP), working on a four-year project in collaboration with BT Research. My research develops AI methods for ultra fine-grained urban land use mapping. I fuse high-resolution satellite, aerial and street-level imagery with LiDAR to identify surface materials, infrastructure assets and vegetation, and so support smarter, lower-disruption network planning. Before joining Surrey, I studied Computer Science at the University of Bath. My background spans machine learning research, including latent reasoning and reinforcement learning for large language models, and full-stack software engineering.
I am a PhD researcher in the Centre for Vision, Speech and Signal Processing (CVSSP), working on a four-year project in collaboration with BT Research. My research develops AI methods for ultra fine-grained urban land use mapping. I fuse high-resolution satellite, aerial and street-level imagery with LiDAR to identify surface materials, infrastructure assets and vegetation, and so support smarter, lower-disruption network planning. Before joining Surrey, I studied Computer Science at the University of Bath. My background spans machine learning research, including latent reasoning and reinforcement learning for large language models, and full-stack software engineering.
Publications
Highlights
- Co-authored "Explain to Gain: Introspective Reinforcement Learning for Enhanced Performance", presented at the World Conference on Explainable Artificial Intelligence (xAI 2025) in Istanbul and published by Springer.
- Co-authored "Explain to Gain: Optimising Performance Through Explainable Reinforcement Learning Parameter Investigation", published by Springer from the International Joint Conference on Computational Intelligence (IJCCI 2025).
Explainable reinforcement learning is usually used to make an agent's decisions easier to understand. This paper uses it to improve training instead. Building on the IxDRL introspection framework, we feed explainability metrics into model-free training loops and adjust each algorithm's exploration in real time: entropy for PPO and epsilon for DQN. Experiments on the CartPole and Taxi environments, covering both on-policy and off-policy learning, show significant performance gains. They show that explainability can improve agent performance as well as make it more transparent.
This paper extends our earlier introspective explainable reinforcement learning work. We add a further explainability metric to the parameters that can be tuned when configuring model-free training. Training parameters are then adjusted dynamically using real-time feedback from this wider set of metrics. Experiments on the CartPole and Taxi environments, with both on-policy and off-policy algorithms, show further significant performance gains. They suggest that deeper explainability can support more finely tuned, reliable and transparent reinforcement learning systems.