Machine learning researchers build a better world model

Researchers have proposed a method that could help robots predict what would happen under actions they have not tried before. The approach uses the gap between a model’s prediction and what actually happens to infer hidden factors, potentially improving reasoning and future predictions without training on every possible outcome.

The preprint, by Randall Balestriero, Kerda Varaku and Yann LeCun, describes “counterfactual world models”, which can be applied to models that predict an average next state. In a river example, if a boat ends up two metres downstream of its predicted position after turning left, the method carries that inferred current into a prediction for turning right. The researchers say the error from this rule shrinks with the square of the time step, and that this rate cannot be improved in the worst case using the same ordinary data.