Polymatrix Competitive Gradient Descent for Reinforcement Learning
Author(s)
Beard, William
Advisor(s)
Editor(s)
Collections
Supplementary to:
Permanent Link
Abstract
We explore the application of a new optimizer for reinforcement learning tasks: Polymatrix Competitive Gradient Descent (PCGD). We discuss its game-theoretic foundation and explore tasks for which it outperforms traditional optimizers by accounting for player interactions. Its implementation is described in detail along with verification steps to ensure the optimizer works as expected. Finally we provide numerical experiments which show the performance of PCGD is comparable to that of Simultaneous Gradient Descent (SimGD) for tasks from the OpenAI multiagent particle environments.
Sponsor
Date
Extent
Resource Type
Text
Resource Subtype
Undergraduate Research Option Thesis