In this work, we investigate, how multistep lookahead affects critical points of Residual Gradient algorithms. We set up a compound Bellman Operator for k consecutive transitions similar to TD(λ) methods and analyse the critical points of the associated Mean Squared Bellman Error (MSBE). By collecting per state multiple successors at once, one can create a more informative objective without increasing the requirements for function approximation architectures. In an empirical analysis, we observe that if one uses Hessian based optimisation to minimise the MSBE, it is not possible to benefit from larger lookahead. Already high convergence speeds and overall lower final error of a Gauss Newton algorithm seem to prevent further improvements by larger lookahead. Only first order gradient descent shows a significant boost in convergence for larger k, emphasizing the importance of multiple steps for existing and successful Deep Reinforcement Learning algorithms. Our results suggest that there are still open questions for Neural Network training in Reinforcement Learning applications.     «

In this work, we investigate, how multistep lookahead affects critical points of Residual Gradient algorithms. We set up a compound Bellman Operator for k consecutive transitions similar to TD(λ) methods and analyse the critical points of the associated Mean Squared Bellman Error (MSBE). By collecting per state multiple successors at once, one can create a more informative objective without increasing the requirements for function approximation architectures. In an empirical analysis, we observe...     »