| Summary: | Advances in reinforcement learning have allowed agents to learn a variety of board games and video games at superhuman levels. Unlike humans - which can generalize to a wide range of tasks with very little experience - these algorithms typically need vast number of experience replays to perform at the same level. In this thesis, we propose a model-based reinforcement learning approach that represents the environment using an explicit symbolic model in the form of a domain-specific language (DSL) that represents the world as a set of discrete objects with underlying latent properties that govern their dynamical interactions. We present a novel, neurally guided, on-line inference technique to recover the structured world representation from raw video observations, with the intent to be used for downstream model-based planning. We qualitatively evaluate our inference performance on classical Atari games, as well as on physics-based mobile games.
|
|---|