Timezone: »

Co-Adaptation of Algorithmic and Implementational Innovations in Inference-based Deep Reinforcement Learning
Hiroki Furuta · Tadashi Kozuno · Tatsuya Matsushima · Yutaka Matsuo · Shixiang (Shane) Gu

Tue Dec 07 04:30 PM -- 06:00 PM (PST) @ Virtual #None

Recently many algorithms were devised for reinforcement learning (RL) with function approximation. While they have clear algorithmic distinctions, they also have many implementation differences that are algorithm-independent and sometimes under-emphasized. Such mixing of algorithmic novelty and implementation craftsmanship makes rigorous analyses of the sources of performance improvements across algorithms difficult. In this work, we focus on a series of off-policy inference-based actor-critic algorithms -- MPO, AWR, and SAC -- to decouple their algorithmic innovations and implementation decisions. We present unified derivations through a single control-as-inference objective, where we can categorize each algorithm as based on either Expectation-Maximization (EM) or direct Kullback-Leibler (KL) divergence minimization and treat the rest of specifications as implementation details. We performed extensive ablation studies, and identified substantial performance drops whenever implementation details are mismatched for algorithmic choices. These results show which implementation or code details are co-adapted and co-evolved with algorithms, and which are transferable across algorithms: as examples, we identified that tanh Gaussian policy and network sizes are highly adapted to algorithmic types, while layer normalization and ELU are critical for MPO's performances but also transfer to noticeable gains in SAC. We hope our work can inspire future work to further demystify sources of performance improvements across multiple algorithms and allow researchers to build on one another's both algorithmic and implementational innovations.

Author Information

Hiroki Furuta (The University of Tokyo)
Tadashi Kozuno (University of Alberta)

Tadashi Kozuno is a postdoc at the University of Alberta. He obtained bachelor and master degrees on neuroscience from Osaka university, and a PhD degree from Okinawa Inst. of Sci. and Tech. His main interest lies in efficient decision making from both theoretical and biological sides.

Tatsuya Matsushima (The University of Tokyo)
Yutaka Matsuo (University of Tokyo)
Shixiang (Shane) Gu (Google Brain, University of Cambridge)

More from the Same Authors