All Projects → DongjunLee → dqn-tensorflow

DongjunLee / dqn-tensorflow

Licence: other
Deep Q Network implements by Tensorflow

Programming Languages

python
139335 projects - #7 most used programming language

Projects that are alternatives of or similar to dqn-tensorflow

Keras Rl2
Reinforcement learning with tensorflow 2 keras
Stars: ✭ 134 (+436%)
Mutual labels:  deep, dqn
Protobuf-Dreamer
A tiled DeepDream project for creating any size of image, on both CPU and GPU
Stars: ✭ 39 (+56%)
Mutual labels:  deep
Differentia.js
No longer being supported or maintained. A Graph Theory & Data Structure Library for JavaScript.
Stars: ✭ 13 (-48%)
Mutual labels:  deep
DQN-Atari
Deep Q-Learning (DQN) implementation for Atari pong.
Stars: ✭ 53 (+112%)
Mutual labels:  dqn
IJCAI2018 SSDH
Semantic Structure-based Unsupervised Deep Hashing IJCAI2018
Stars: ✭ 38 (+52%)
Mutual labels:  deep
pytorch-rl
Pytorch Implementation of RL algorithms
Stars: ✭ 15 (-40%)
Mutual labels:  dqn
iextrading4j-hist
IEX Trading library to parse TOPS and DEEP multicast packets
Stars: ✭ 20 (-20%)
Mutual labels:  deep
Explorer
Explorer is a PyTorch reinforcement learning framework for exploring new ideas.
Stars: ✭ 54 (+116%)
Mutual labels:  dqn
CarND-VehicleDetection
vehicle detection with deep learning
Stars: ✭ 34 (+36%)
Mutual labels:  deep
distributedRL
A framework for easy prototyping of distributed reinforcement learning algorithms
Stars: ✭ 93 (+272%)
Mutual labels:  dqn
cca zoo
Canonical Correlation Analysis Zoo: A collection of Regularized, Deep Learning based, Kernel, and Probabilistic methods in a scikit-learn style framework
Stars: ✭ 103 (+312%)
Mutual labels:  deep
DQN-pytorch
A PyTorch implementation of Human-Level Control through Deep Reinforcement Learning
Stars: ✭ 23 (-8%)
Mutual labels:  dqn
introspected
Introspection for serializable arrays and JSON friendly objects.
Stars: ✭ 75 (+200%)
Mutual labels:  deep
defaults-deep
Like `extend` but recursively copies only the missing properties/values to the target object.
Stars: ✭ 26 (+4%)
Mutual labels:  deep
map-keys-deep-lodash
Map/rename keys recursively with Lodash
Stars: ✭ 16 (-36%)
Mutual labels:  deep
AI BIG DATAS ALGORITHM
大数据+人工智能+数据结构相关案例项目
Stars: ✭ 28 (+12%)
Mutual labels:  dqn
jest-expect-contain-deep
Assert deeply nested values in Jest
Stars: ✭ 68 (+172%)
Mutual labels:  deep
rl
Reinforcement learning algorithms implemented using Keras and OpenAI Gym
Stars: ✭ 14 (-44%)
Mutual labels:  deep
Tensorflow-Wide-Deep-Local-Prediction
This project demonstrates how to run and save predictions locally using exported tensorflow estimator model
Stars: ✭ 28 (+12%)
Mutual labels:  deep
chainer-notebooks
Jupyter notebooks for Chainer hands-on
Stars: ✭ 23 (-8%)
Mutual labels:  dqn

Deep Q Network

Paper

TO DO

  • Test: Atari
    • more complex ConvNet model
  • use TensorBoard
    • average loss
    • average q
    • average reward (consecutive 100 episode)
    • episode reward

Config

python main.py -h

  --discount_rate DISCOUNT_RATE
                        Initial discount rate.
  --replay_memory_length REPLAY_MEMORY_LENGTH
                        Number of replay memory episode.
  --target_update_count TARGET_UPDATE_COUNT
                        DQN Target Network update count.
  --max_episode_count MAX_EPISODE_COUNT
                        Number of maximum episodes.
  --batch_size BATCH_SIZE
                        Batch size. (Must divide evenly into the dataset
                        sizes)
  --frame_size FRAME_SIZE
                        Frame size. (Stack env's observation T-n ~ T)
  --model_name MODEL_NAME
                        DeepLearning Network Model name (MLPv1, ConvNetv1)
  --learning_rate LEARNING_RATE
                        Batch size. (Must divide evenly into the dataset
                        sizes)
  --gym_result_dir GYM_RESULT_DIR
                        Directory to put the gym results.
  --gym_env GYM_ENV     Name of Open Gym's enviroment name. (CartPole-v0,
                        CartPole-v1, MountainCar-v0)
  --step_verbose [STEP_VERBOSE]
                        verbose every step count
  --step_verbose_count STEP_VERBOSE_COUNT
                        verbose step count

Model

1. MLPv1

  • hidden layer (16, 64, 32)
  • AdamOptimizer

2. ConvNetv1

  • 3 Conv + MaxPool Layers (kernel_size [3, 3, 3], filters [32, 64, 128])
  • 2 Fully Connected Layers (hidden_size [128, 32])
  • AdamOptimizer

3. ConvNetv2

  • 5 Conv + MaxPool Layers (kernel_size [7, 5, 3, 3, 3], filters [126, 256, 512, 512, 512]
  • 2 Fully Connected Layers (hidden_size [1024, 256])
  • AdamOptimizer

Expertiments

Classic control

CartPole-v0 CartPole-v1 MountainCar-v0
defines "solving" as getting average reward of 195.0 over 100 consecutive trials. defines "solving" as getting average reward of 475.0 over 100 consecutive trials. defines "solving" as getting average reward of -110.0 over 100 consecutive trials.
Model : MLPv1 Model : MLPv1 Model : MLPv1
Clear : after 177 episode Clear : after 791 episode Clear : after 1182 episode
images images images

Atari

Assault-ram-v0
Maximize your score
Model : ConvNetv2
Score : 421.12 (average from 100 consecutive trials)
images
2000 Episode (Learn something.. but still stupid)
Breakout-ram-v0
Maximize your score
Model : ConvNetv1
Score : 9.69 (average from 100 consecutive trials)

Reference

Note that the project description data, including the texts, logos, images, and/or trademarks, for each open source project belongs to its rightful owner. If you wish to add or remove any projects, please contact us at [email protected].