Timezone: »

 
A functional mirror ascent view of policy gradient methods with function approximation
Sharan Vaswani · Olivier Bachem · Simone Totaro · Matthieu Geist · Marlos C. Machado · Pablo Samuel Castro · Nicolas Le Roux

We use functional mirror ascent to propose a general framework (referred to as FMA-PG) for designing policy gradient methods. The functional perspective distinguishes between a policy's functional representation (what are its sufficient statistics) and its parameterization (how are these statistics represented), and naturally results in computationally efficient off-policy updates. For simple policy parameterizations, the FMA-PG framework ensures that the optimal policy is a fixed point of the updates. It also allows us to handle complex policy parameterizations (e.g., neural networks) while guaranteeing policy improvement. Our framework unifies several PG methods and opens the way for designing sample-efficient variants of existing methods. Moreover, it recovers important implementation heuristics (e.g., using forward vs reverse KL divergence) in a principled way. With a softmax functional representation, FMA-PG results in a variant of TRPO with additional desirable properties. It also suggests an improved variant of PPO, whose robustness we empirically demonstrate on MuJoCo.

Author Information

Sharan Vaswani (University of Alberta)
Olivier Bachem (Google Brain)
Simone Totaro (Mila)
Matthieu Geist (Google)
Marlos C. Machado (DeepMind)
Pablo Samuel Castro (Google Brain)

Pablo was born and raised in Quito, Ecuador, and moved to Montreal after high school to study at McGill. He stayed in Montreal for the next 10 years, finished his bachelors, worked at a flight simulator company, and then eventually obtained his masters and PhD at McGill, focusing on Reinforcement Learning. After his PhD Pablo did a 10-month postdoc in Paris before moving to Pittsburgh to join Google. He has worked at Google for almost 6 years, and is currently a research Software Engineer in Google Brain in Montreal, focusing on fundamental Reinforcement Learning research, as well as Machine Learning and Music. Aside from his interest in coding/AI/math, Pablo is an active musician (https://www.psctrio.com), loves running (5 marathons so far, including Boston!), and discussing politics and activism.

Nicolas Le Roux (Google)

More from the Same Authors