contra_3_rl

• Reinforcement Learning attempts to beat Contra 3 for the SNES
• Run using Retro Gym (https://github.com/openai/retro)
• Levels 1-6 cleared using OpenAI Baselines PPO implementation

To run levels 1-6 easy clears:

• Install Open AI Baselines (https://github.com/openai/baselines). These runs use the November 11th, 2018 version of PPO with some slight modifications (see the misc directory)
• Place Contra III-Snes directory in corresponding Retro Gym directory and modify scenario.json and data.json if necessary (see readme.md)
• run ppo2_contra_baselines_agent (...).py from the command line

Explanation of directories:

• /log: stores TensorBoard data for successful level clears
• /logs: stores .csv files with basic stats like reward and timesteps
• /videos: videos of cleared runs
• /ppo_save: saved models

Explantion of files:

• ppo2_contra_baselines_agent (...).py: launches the runs
• render.py: view videos of episodes
• retro_movie_debugger.py: view videos of episodes and examing env outputs like rewards, info, lua script print messages etc. Useful for debugging
• monitor_graphs.ipynb: Jupyter notebook that shows data from the /logs files to see results
• image_test.ipynb: visual testing of observation space so you can see what your agent sees

Command Line Arguments for Level 1-6 Clears:

• python3 ppo2_contra_medium.py --game ContraIII-Snes --state level4.1player.hard.100lives --num_env 21 --seed 28 --scenario_number 1 --stochastic_frame_skip 4 --scale_reward 0.01 --skip_prob 0.25 --scenario scenario_medium_lvl1_run0 --time_limit 20000 --network contra_mixed_cnn
• python3 ppo2_contra_medium.py --game ContraIII-Snes --state level3.1player.medium.100lives --num_env 21 --seed 644 --scenario_number 1 --stochastic_frame_skip 4 --scale_reward 0.01 --skip_prob 0.25 --scenario scenario_medium_lvl6_run0 --time_limit 20000 --network contra_mixed_cnn
• python3 ppo2_contra_medium.py --game ContraIII-Snes --state level1.1player.hard.100lives --num_env 21 --seed 203 --scenario_number 1 --stochastic_frame_skip 4 --scale_reward 0.01 --skip_prob 0.25 --scenario scenario_medium_lvl1_run0 --time_limit 20000 --network contra_mixed_cnn
• python3 ppo2_contra_medium.py --game ContraIII-Snes --state level1.1player.hard.100lives --num_env 21 --seed 7262 --scenario_number 1 --stochastic_frame_skip 4 --scale_reward 0.01 --skip_prob 0.25 --scenario scenario_medium_lvl1_run0 --time_limit 20000 --network cnn
• python3 ppo2_contra_medium.py --game ContraIII-Snes --state level1.1player.medium.100lives --num_env 21 --seed 5242 --scenario_number 1 --stochastic_frame_skip 4 --scale_reward 0.01 --skip_prob 0.25 --scenario scenario_medium_lvl1_run0 --time_limit 20000 --network cnn
• python3 ppo2_contra_baselines_agent.py --game ContraIII-Snes --state level1.1player.easy.100lives --num_env 21 --seed 112 --scenario_number 3 --stochastic_frame_skip 4 --scale_reward 0.01 --skip_prob 0.25 --scenario scenario16
• python3 ppo2_contra_baselines_agent.py --game ContraIII-Snes --state level2.1player.easy.100lives --num_env 21 --seed 443 --scenario_number 7 --stochastic_frame_skip 4 --scale_reward 0.01 --skip_prob 0.25 --scenario scenario_lvl2_run2
• python3 ppo2_contra_baselines_agent.py --game ContraIII-Snes --state level3.1player.easy.100lives --num_env 21 --seed 279 --scenario_number 1 --stochastic_frame_skip 4 --scale_reward 0.01 --skip_prob 0.25 --scenario scenario_lvl3_run12 --time_limit 20000
• python3 ppo2_contra_baselines_agent.py --game ContraIII-Snes --state level4.1player.easy.100lives --num_env 22 --seed 11143 --scenario_number 2 --stochastic_frame_skip 4 --scale_reward 0.01 --skip_prob 0.25 --scenario scenario22
• python3 ppo2_contra_baselines_agent.py --game ContraIII-Snes --state level5.1player.easy.100lives --num_env 21 --seed 46 --scenario_number 1 --stochastic_frame_skip 4 --scale_reward 0.01 --skip_prob 0.25 --scenario scenario_lvl5_run6
• python3 ppo2_contra_baselines_agent.py --game ContraIII-Snes --state level6.1player.easy.100lives --num_env 21 --seed 222372 --scenario_number 1 --stochastic_frame_skip 4 --scale_reward 0.01 --skip_prob 0.25 --scenario scenario18