%0 Journal Article %1 hiraoka2020metamodelbased %A Hiraoka, Takuya %A Imagawa, Takahisa %A Tangkaratt, Voot %A Osa, Takayuki %A Onishi, Takashi %A Tsuruoka, Yoshimasa %D 2020 %K meta-learning optimization reinforcement-learning sampling %T Meta-Model-Based Meta-Policy Optimization %U http://arxiv.org/abs/2006.02608 %X Model-based reinforcement learning (MBRL) has been applied to meta-learning settings and demonstrated its high sample efficiency. However, in previous MBRL for meta-learning settings, policies are optimized via rollouts that fully rely on a predictive model for an environment, and thus its performance in a real environment tends to degrade when the predictive model is inaccurate. In this paper, we prove that the performance degradation can be suppressed by using branched meta-rollouts. Based on this theoretical analysis, we propose meta-model-based meta-policy optimization (M3PO), in which the branched meta-rollouts are used for policy optimization. We demonstrate that M3PO outperforms existing meta reinforcement learning methods in continuous-control benchmarks.

@article{hiraoka2020metamodelbased, abstract = {Model-based reinforcement learning (MBRL) has been applied to meta-learning settings and demonstrated its high sample efficiency. However, in previous MBRL for meta-learning settings, policies are optimized via rollouts that fully rely on a predictive model for an environment, and thus its performance in a real environment tends to degrade when the predictive model is inaccurate. In this paper, we prove that the performance degradation can be suppressed by using branched meta-rollouts. Based on this theoretical analysis, we propose meta-model-based meta-policy optimization (M3PO), in which the branched meta-rollouts are used for policy optimization. We demonstrate that M3PO outperforms existing meta reinforcement learning methods in continuous-control benchmarks.}, added-at = {2020-06-05T11:16:20.000+0200}, author = {Hiraoka, Takuya and Imagawa, Takahisa and Tangkaratt, Voot and Osa, Takayuki and Onishi, Takashi and Tsuruoka, Yoshimasa}, biburl = {https://www.bibsonomy.org/bibtex/2a162d2fdb84fc338378efad1ba8a84f7/kirk86}, description = {[2006.02608] Meta-Model-Based Meta-Policy Optimization}, interhash = {1bb76688b7041c0e6355c65e119a75b7}, intrahash = {a162d2fdb84fc338378efad1ba8a84f7}, keywords = {meta-learning optimization reinforcement-learning sampling}, note = {cite arxiv:2006.02608}, timestamp = {2020-06-05T11:16:20.000+0200}, title = {Meta-Model-Based Meta-Policy Optimization}, url = {http://arxiv.org/abs/2006.02608}, year = 2020 }