Skip to content

Latest commit

 

History

History
57 lines (40 loc) · 2.29 KB

README.md

File metadata and controls

57 lines (40 loc) · 2.29 KB
Raw

DRL Dodgeball

DRL (Deep Reinforcement Learning) has led to the creation of agents that exhibit fascinatingly complex and intelligent behavior, especially in the area of computer games like Dota 2 and StarCraft II. An exciting area of research for is to develop digital agents that eventually get deployed into physical robots, a task which OpenAI’s Dactyl project demonstrates requires high-fidelity training environments. In this project, we designed a physically realistic simulation with high-dimensional sensory data sources, and trained an agent using a refined deep Q-network in it.

Files

train.py

  • Main file used for training, currently training agents with refined DQN
  • Command: "py train.py"

train_DQN.py

  • Training agents under the Unity environment with refined DQN
  • Command: "py train_DQN.py"

train_DoubleDQN.py

  • Training agents under the Unity environment with Double DQN
  • Command: "py train_DoubleDQN.py"

train_vanilla.py

  • Training agents under the Unity environment with vanilla Q Learning
  • Command : "py train_vanilla.py"

utils.py

  • Modularized utility function which contains abstract functions necessary for training our agent

experience_replay.py

  • Container for the storage and use of memories/experiences according to our prioritized replay buffer.

networkparams.pt

  • Configuration file for defining network parameters and topology

networks.py

  • Pytorch network definition following the parameters from networkparams.pt

PolicyGradients.ipynb

  • Attempt at using Policy Gradients as an alternative learning mechanism to train intelligent agents on our environment

QLearning.ipynb

  • Generalized Q Learning to solve MDP problems

QLearningAgentAndy.py

  • Implementation of generic Q learning on basic example environments from Unity

\runs

  • TensorBoard based results of training runs

\environments

  • Built Unity environments

Future Work

  • Accomodate continuous action spaces
  • Acommodate multi-dimensional action spaces
  • Accomodate visual observations (first-person and birds-eye)
  • Experiment with more robust neural network architectures
  • Experiment with different optimizers

poster