An operator view of policy gradient methods
Ghosh, Dibya, Machado, Marlos C., Roux, Nicolas Le
–arXiv.org Artificial Intelligence
We cast policy gradient methods as the repeated application of two operators: a policy improvement operator $\mathcal{I}$, which maps any policy $\pi$ to a better one $\mathcal{I}\pi$, and a projection operator $\mathcal{P}$, which finds the best approximation of $\mathcal{I}\pi$ in the set of realizable policies. We use this framework to introduce operator-based versions of traditional policy gradient methods such as REINFORCE and PPO, which leads to a better understanding of their original counterparts. We also use the understanding we develop of the role of $\mathcal{I}$ and $\mathcal{P}$ to propose a new global lower bound of the expected return. This new perspective allows us to further bridge the gap between policy-based and value-based methods, showing how REINFORCE and the Bellman optimality operator, for example, can be seen as two sides of the same coin.
arXiv.org Artificial Intelligence
Jun-22-2020
- Country:
- North America
- United States > New Jersey
- Mercer County > Princeton (0.04)
- Canada > Quebec
- Montreal (0.14)
- United States > New Jersey
- North America
- Genre:
- Research Report (0.40)
- Technology: