Telmo de Menezes e Silva Filho

Agents can learn reusable, high-level causal abstractions from raw experience with orders-of-magnitude greater sample efficiency than neural network baselines, thanks to a new framework combining online model learning, repair, and predicate invention.