Value equivalence prediction Sample Clauses

Value equivalence prediction. Finally, an intermediate approach between model based and model free methods is offered by value iteration [B59]. Here generated trajectories are only constrained to the real environment in the sense that they should have the same cumulative reward. They have proven to be effective in high dimensional observation spaces. [B60] To summarize, work from [B49] shows that predictive models generalize well enough to overcome their implicit bias but suffer from compounding errors when making long horizon rollouts. Like PETS [B39] they use probabilistic model ensembles combined with a stable off-policy model free optimizer. [B61]