# Trust Region Policy Optimization

**Authors:** John Schulman, Sergey Levine, Philipp Moritz, Michael I. Jordan, Pieter Abbeel  
**arXiv:** [1502.05477](https://arxiv.org/abs/1502.05477)  
**Date:** 2015-02-19  
**Summary page:** https://www.summarizepaper.com/en/paper/1502.05477v1/

## Key points

- "Trust Region Policy Optimization" paper by John Schulman, Sergey Levine, Philipp Moritz, Michael I. Jordan, and Pieter Abbeel introduces TRPO algorithms for learning control policies
- TRPO offers a promising approach for solving complex tasks in diverse environments
- Key contribution: development of a policy update scheme ensuring monotonic improvement for large-scale problems
- Authors conducted experiments on challenging tasks like simulated robotic locomotion and playing Atari games using images as input
- Showcases potential of reinforcement learning techniques in various domains

## Summary

The paper "Trust Region Policy Optimization" by John Schulman, Sergey Levine, Philipp Moritz, Michael I. Jordan, and Pieter Abbeel introduces a family of algorithms designed for learning control policies. These algorithms are called trust region policy optimization (TRPO) and offer a promising approach to solving complex tasks in diverse environments. The key contribution of this work is the development of a policy update scheme that guarantees monotonic improvement and can be applied to large-scale problems. The authors also conduct experiments on challenging tasks such as simulated robotic locomotion and playing Atari games using images as input. Overall, their work showcases the potential for using reinforcement learning techniques in various domains.