DQN Experiment

Introduction

The experiment is to find out the ways to improve simple DQN performance.

Installation

1. Install DeepMind Lab
2. Replace BUILD in DeepMind Lab repo, ensure it has the correct path to srcs
3. Copy agent_*.py scripts to lab/python/
4. Make the necessary adjustments to the .lua file to configure a different map.

Train a Model

The default path to save a model is 'models/'

gym$ bazel run :DQN_imgstk -- --level_script=/home/pengk/git_ws/lab/game_scripts/levels/seekavoid_arena_01.lua   --num_steps 1000000

Run a Pre-trained Model

gym$ bazel run :DQN_imgstk -- --level_script=/home/pengk/git_ws/lab/game_scripts/levels/seekavoid_arena_01.lua   --num_steps 3050 --Test True   --model  /home/pengk/git_ws/lab/models/lenet_4img_1000000.pt

Result

Reward:

Apple: +1 | Lemon: -1