DQN

This is a hobby project to learn DQN reinforcement learning.

The implementation is quite standard DQN, including

• Simple feedforward network to approximate state-value function
• Replay memory
• Separate target network that is "smoothed" version of policy net

I'm planning to add some experimental features. Now there are few

• Option for ensemble learning
  • Voting best action based on multiple models
  • Finding most uncertain actions in exploration phase using multiple models
• Option for adding reward bonus based on (hashed) state count as exploration strategy
• Option to use priority based sampling from replay memory; priority defined by temporal difference error

Examples

Grid world with obstacles

Game

• Grid world, 30 x 30
• One target and 15 obstacles
• Rewards:
  • target: +10
  • obstacle: -10
  • otherwise: -0.01
• Robot starts from (0,0) and has 200 steps time to find target at (20, 20)

Testing

• Test done with for different methods:
  • Standard
  • Priority sampling, based on temporal difference errors
  • Count based exploration
  • Ensemble
• For each method, the run is repeated 20 times
• Each run is 1000 episodes
• Smoothed (running mean, 50 episodes) reward is calculated for each run

See details: src/samples/sample_method_comparison.py

Results

Example run (method: priority sampling):

20 runs with standard DQN:

20 runs with priority sampling:

20 runs with count based exploration:

20 runs with ensemble:

Comparison of methods, average of 20 runs from each:

References

Prioritized Experience Replay

Tom Schaul, John Quan, Ioannis Antonoglou, David Silver

http://arxiv.org/abs/1511.05952

#Exploration: A Study of Count-Based Exploration for Deep Reinforcement Learning

Haoran Tang, Rein Houthooft, Davis Foote, Adam Stooke, Xi Chen, Yan Duan, John Schulman, Filip De Turck, Pieter Abbeel

https://arxiv.org/abs/1611.04717