The experiment is to find out the ways to improve simple DQN performance.
- Install DeepMind Lab
- Replace
BUILD
in DeepMind Lab repo, ensure it has the correct path to srcs - Copy agent_*.py scripts to lab/python/
- Make the necessary adjustments to the
.lua
file to configure a different map.
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
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
Reward:
Apple: +1 | Lemon: -1