Agent-DOoM

This is a RL Agent that learns to play Doom Deadly Corridor. The Agent is trained using a Double Deep Q-Learning with Duel Architecture and Prioritized Experience Replay.

Requirements

• VizDoom (Doom's Enviornment)
• PyTorch
• OpenCV

Contents

• utils.py : Prioritized Memory Replay , Sum Trees , Preprocessing classes and fucntions.
• model.py : Dueling Deep Q Network
• doom.py : Training and Testing Agent.
• dqnet.py : weights after 1 hour training on GPU
• deadly_corridor.cfg : Enviornment Configuration File
• out.avi : Output video (Agent learnt to escape from enemies)

Usage

• To Train from starting : python3 doom.py --train=True --gpu=True --save_weights=True
• Continue Training : python3 doom.py --weights=./dqnet.pt --train=True --gpu=True --save_weights=True

Flags

Flag Name	Description	Default
--learning_rate	Learning rate for DQN	0.025
--episodes	Number of episodes for training	5000
--steps	Maximum steps per episode	5000
--tau	Time to refil target network weights	3000
--explore_stop	Stopping exploration probability	0.01
--decay_rate	Decay rate for exploration probability	0.00005
--discount_factor	Discounting factor for returns	0.95
--pretrain_length	number of iterations to fill replay memory	10000
--mem_size	Size of the replay memory	10000
--train	want to train or not	False
--batch_size	Batch size to sample from replay memory	64
--gpu	Use GPU	False
--weights	load pretrained weights	None
--save_weights	save the DQN weights	False
--graph	Draw summary graph	False
--test	Generate test video	False

Note 1 : Default hyper-paramemters are used for training

Note 2: This Agent is trained only for 1 Hour on GPU and it has learned to escape from two enemies . Further training is needed to reach optimal policy