This paper studies a gradient temporal difference (GTD) algorithm using neural network (NN) function approximators to minimize the mean squared Bellman error (MSBE). For off-policy learning, we show that the minimum MSBE problem can be recast into a min-max optimization involving a pair of over-parameterized primal-dual NNs. The resultant formulation can then be tackled using a neural GTD algorithm. We analyze the convergence of the proposed algorithm with a 2-layer ReLU NN architecture using m neurons and prove that it computes an approximate optimal solution to the minimum MSBE problem as m ! 1.
|Original language||English (US)|
|Journal||Advances in Neural Information Processing Systems|
|State||Published - 2020|
|Event||34th Conference on Neural Information Processing Systems, NeurIPS 2020 - Virtual, Online|
Duration: Dec 6 2020 → Dec 12 2020
Bibliographical noteFunding Information:
Acknowledgement & Funding Disclosure The authors would like to thank Mr. Alan Lun (CUHK) for conducting the preliminary numerical experiments in this paper. H.-T. Wai is supported by the CUHK Direct Grant #4055113. M. Hong is supported in part by NSF under Grant CCF-1651825, CMMI-172775, CIF-1910385 and by AFOSR under grant 19RT0424.
© 2020 Neural information processing systems foundation. All rights reserved.