research

Learning representations for reinforcement learning (RL) has shown much promise for continuous control. In this project, we investigate using vector quantization to prevent representation collapse when learning representations for RL using a self-supervised latent-state consistency loss.

Sequential learning paradigms pose challenges for gradient-based deep learning due to difficulties incorporating new data and retaining prior knowledge. While Gaussian processes elegantly tackle these problems, they struggle with scalability and handling rich inputs, such as images.

This project seeks to evaluate and compare different approaches for learning dynamics models in model-based RL. In particular, we plan to compare different approximate inference techniques (e.g. Laplace approximation, MC dropout, variational inference), as well as ensemble methods, to understand why they either succeed or fail in different environments.

This work presents a learning-based control method for navigating to a target state in unknown, or partially unknown, multimodal dynamical systems. In particular, it develops a model-based reinforcement learning algorithm that can remain in a desired dynamics mode with high probability. For example, if some of the dynamics modes are believed to be inoperable.

This work presents a two-stage method to perform trajectory optimisation in multimodal dynamical systems with unknown nonlinear stochastic transition dynamics. The method finds trajectories that remain in a preferred dynamics mode where possible and in regions of the transition dynamics model that have been observed and can be predicted confidently.

This work introduces a variational lower bound for the Mixture of Gaussian Process Experts model with a GP-based gating network based on sparse GPs. The model (and inference) are implemented in GPflow/TensorFlow.