A New Representation of Successor Features for Transfer across Dissimilar Environments

Transfer in reinforcement learning is usually achieved through generalisation\nacross tasks. Whilst many studies have investigated transferring knowledge when\nthe reward function changes, they have assumed that the dynamics of the\nenvironments remain consistent. Many real-world RL problems require transfer\namong environments with different dynamics. To address this problem, we propose\nan approach based on successor features in which we model successor feature\nfunctions with Gaussian Processes permitting the source successor features to\nbe treated as noisy measurements of the target successor feature function. Our\ntheoretical analysis proves the convergence of this approach as well as the\nbounded error on modelling successor feature functions with Gaussian Processes\nin environments with both different dynamics and rewards. We demonstrate our\nmethod on benchmark datasets and show that it outperforms current baselines.\n