Reviews: Active Exploration for Learning Symbolic Representations

This is a very interesting paper, with multiple complementary ideas. It advocates model-based active exploration (model learning seeking regions of uncertainty). Instead of doing this in raw state space, it proposes a method for abstracting states to symbols based on factoring and clustering the state space. The exploration is then done by MCTS-planning in a (sampled) symbolic model. The task setup evaluates pure exploration (ignoring all rewards) on a two different domains.