Deep deterministic policy gradient
Download as PPTX, PDF1 like826 views
This document discusses Deep Deterministic Policy Gradient (DDPG), a reinforcement learning algorithm for problems with continuous state and action spaces. DDPG uses an actor-critic method with experience replay and soft target updates to learn a policy in an off-policy manner. It demonstrates how DDPG can be used to train an agent to drive a vehicle in a simulator by designing a reward function, but notes that designing effective rewards, avoiding local optima, instability, and data requirements are challenges for DDPG.
Related topics:1 of 19
Downloaded 11 times
Deep deterministic policy gradient
- 1. Deep Deterministic Policy Gradient DDPG
- 2. History
- 3. ML methods
- 4. ML methods
- 8. Unsupervised
- 9. Unsupervised
- 11. Gradient Descent
- 13. Grid worlds
- 14. Value function vs Policy
- 15. Actor critic
- 17. DDPG - Continuous state and action space - Replay buffer - Soft updates - Exploration noise
- 18. Pitfalls - Designing reward function is very hard - Tends to get stuck into local optima - Unstable - Needs lots of training samples