Skip to content

Latest commit

 

History

History
157 lines (137 loc) · 7 KB

README.md

File metadata and controls

157 lines (137 loc) · 7 KB
Raw

ViZDoom Bot

We're ViZDoom Bot team of CSCI527 Spring 2021. This repository contains all the code and part of saved model weights of our team throughout the whole semester. Please visit our website for more details.

Introduction of ViZDoom

The ViZDoom competition was an artificial intelligence competition held from 2016 to 2018 in which artificial agents making use of the ViZDoom platform and having access to only Doom's visual buffer and UI statistics engaged in a series of tests on separate tracks, including death matches and navigation challenges, to determine which agent / AI method was superior at playing the game doom.

The goal of this project will be to use current machine learning techniques to construct an artificial agent capable of competent play at the game Doom in general, and more specifically an agent which would have been capable of successfully competing in the ViZDoom competition and whose performance will (preferably) exceed that of the best previous competitors in the competition.

Project Structure

  • archive_code: Self-implemented agents using TensorFlow v2, which are not used after midterm and became archived since then. Please refer to individual README files within each one of them.
  • archive_models: Saved models for some scenarios. Copy zip files into the folder with the same name located in the repo root to use them.
  • common: Common building blocks of our agent. The final version of our agent is built with the help of stable-baselines. We added some custom networks, functions, and wrappers on top of it.
  • scenarios: All the scenarios used while training our agents.
  • All other folders: Each of them specifies a particular setting (hyper-params, action space, network params, etc.) for the agent that can be then trained and evaluated on a particular scenario.

Dependencies

See requirements.txt for the full list.

WARNING: stable-baselines only support Tensorflow <= 1.15 for now. So be sure to install the right version. In addition, if you plan to use GPU, here's a list of cuda lib versions for your reference:

  • cuda: 10.0.130
  • cudnn: 7.6.5

TIPS: using Conda to manage different lib versions and virtual environments is highly recommended.

Directions to Run the Agent

After installing all the required dependencies, follow the directions below to train & evaluate the agent. Take simpler_basic_lstm as an example:

cd simpler_basic_lstm

# Edit constants.py to change action space, input frame size, game args, etc.

# Edit params.py to tweak parameters and hyper-parameters.

# Run the training process. If interrupted halfway, models will be safely saved.
python3 train_ppo_lstm.py

# Run the evaluate process
python3 evaluate_ppo_lstm.py

# Record a video of the agent playing the game along with visualized output of
# the policy network.
python3 record_evaluate_ppo_lstm.py

Directions to Run a Deathmatch

# In terminal 1
# Start the host
cd flatmap_host
# Edit host_simple.py to adjust number of clients, number of built-in bots,
# and whether the host will be a player or a specular. Then:
python3 host_simple.py

# In terminal 2
# Start agent 1
cd flatmap_agent
python3 deathmatch.py

# (Optional)
# In terminal 3
# Start agent 2
cd flatmap_agent2
python3 deathmatch.py

Scenarios Information

'_lvn' postfix where n ranges from 1 to 9 indicates the difficulty level of built-in bots with varied Speed and Health values. Specifically:

  • lv1: Speed = 0.2, Health = 40. (Pistol only)
  • lv2: Speed = 0.2, Health = 40.
  • lv3: Speed = 0.4, Health = 40.
  • lv4: Speed = 0.4, Health = 60.
  • lv5: Speed = 0.6, Health = 60.
  • lv6: Speed = 0.8, Health = 60.
  • lv7: Speed = 0.8, Health = 80.
  • lv8: Speed = 1.0, Health = 100.
  • lv9: Speed = random(0.2, 1.0), Health = random(40, 100).

'_aug' postfix indicates the agent is trained using AugmentedPPO in which force exploration is enabled with explore_probability = 0.1.

'_larger' postfix indicates the agent uses a larger network instead of the default nature_cnn provided by stable-baselines.

  • cig_map1: The scenario for Track 1a (limited deathmatch).
  • cig_map1_less_penalty: The scenario for Track 1a with less shaping penalties for suicide.
  • cig_map2/3: The scenario for Track 2 (full deathmatch).
  • deadly_corridor: Refer to official docs of ViZDoom.
  • deathmatch: Refer to official docs of ViZDoom.
  • defend_the_center: Refer to official docs of ViZDoom.
  • defend_the_line: Refer to official docs of ViZDoom.
  • flatmap_agent/agent2: Join a deathmatch as a client.
  • flatmap_host: Host a deathmatch as the host.
  • flatmap: Simplified version of cig_map1.
  • flatmap_acc: Penalize more on missing shots in hope of boosting aiming accuracy.
  • health_gathering: Refer to official docs of ViZDoom.
  • midterm_demo_2/3: A scenario where the agent is spawned with scattered pickups and with a few monsters including HellKnights. The goal is to kill all HellKnights within time limit.
  • my_cig_01_xxxx_xxx: Deprecated version of cig_map1. Preliminary tests before midterm.
  • my_way_home: Refer to official docs of ViZDoom.
  • my_way_home_lstm: Agent trained on my_way_home without any curiosity mechanisms.
  • my_way_home_lstm_curiosity: Agent trained on my_way_home with RDN.
  • oblige_0: Agent trained and tested on the #0 of PyOblige maps.
  • oblige_multi: Agent trained and tested on different PyOblige maps.
  • oblige_pretrain_aiming: A scenario where the agent id spawned with a random weapon facing a random monster positioned at a random place. The aim of the scenario is to teach the agent to aim and shoot, providing a better starting point for training on oblige maps afterwards.
  • predict_position: Refer to official docs of ViZDoom.
  • rocket_basic/basic2: Refer to official docs of ViZDoom.
  • rocket_basic_easy: Simpler version of rocket_basic where the monster can be killed in one shot.
  • rocket_basic_easy_random: Randomized version of rocket_basic_easy where the initial weapon and monster type are chosen randomly.
  • rocket_medium: Harder version of rocket_basic_easy where in some cases the agent must move around to pick up the rocket launcher and rockets before shooting at the monster.
  • simpler_basic: Refer to official docs of ViZDoom.
  • simpler_deathmatch: Simpler version of deathmatch with reduced action space.
  • simpler_deathmatch_lstm: simpler_deathmatch with policy network integrating LSTM.
  • simpler_my_way_home: Simpler version of my_way_home where only 4 rooms are in the map instead of 7.
  • simpler_my_way_home_lstm_curiosity: simpler_my_way_home with LSTM + RDN.
  • simpler_my_way_home_lstm_icm: simpler_my_way_home with LSTM + ICM.
  • simpler_my_way_home_lstm_icm2: simpler_my_way_home_icm with curiosity_weight = 10.0.
  • simpler_my_way_home_lstm_icm3: simpler_my_way_home_icm with curiosity_weight = 100.0.
  • simpler_my_way_home_lstm_icm4: simpler_my_way_home_icm with curiosity_weight = 5.0.
  • take_cover: Refer to official docs of ViZDoom.