LIS (Life in Silico)
=============
Algorithm
Algorithm Reference
-
Mnih, V. et al. Human-level control through deep reinforcement learning. Nature 518, 529–533 (2015)
-
http://www.nature.com/nature/journal/v518/n7540/abs/nature14236.html
-
A. Krizhevsky, I. Sutskever, and G. Hinton. ImageNet classification with deep convolutional neural networks. In NIPS, 2012.
Requirements
- python 2.7
Install
Ubuntu
Install Unity experimental-build version:
wget http://download.unity3d.com/download_unity/linux/unity-editor-installer-5.3.4f1+20160317.sh
sudo sh unity-editor-installer-5.3.4f1+20160317.sh
# run Unity
./unity-editor-5.3.4f1/Editor/Unity
# if background is pink, install:
sudo apt-get install lib32stdc++6 -y
install python modules:
pip install -r python-agent/requirements.txt
Mac
Install Unity.
install python modules:
pip install -r python-agent/requirements.txt
Windows
Building simulator on Windows10 (Japanese)
Quick Start
download data:
./fetch.sh
Next, run python module as a server.
cd python-agent
python server.py
Open unity-sample-environment with Unity and load Scenes/Sample.
Press Start Buttn. This will take a few minuts for loading caffe model.
You can watch reward history:
cd python-agent
python plot_reward_log.py
This graph is a "sample" scene result. It takes about 6 hours on GPU Machine.
SampleLikesAndDislikes scene result movie
Multi Agent
This is supported only SYNC mode. ASYNC mode is not supprted.
Start multi agent server:
cd python-agent
python multi_agent.py --agent-count=2
Next, open unity-sample-environment and load Scenes/SampleMultiAgent.
You can watch reward history:
python plot_reward_log.py --log-file=reward_0.log
System Configuration
- Client: Unity
- Server: python module
- Communication: Socket (WebSocket over TCP) using MessagePack
Tips
Simulate faster
Select "SceneController" in Hierarchy tab and change "Time Scale".
This will make simulation more faster, but it will be slow gui response.
Module Reference
-
MessagePack for Unity
-
Copyright (C) 2011-2012 Kazuki Oikawa, Kazunari Kida
-
Apache License, Version 2.0
-
Assets/Packages/msgpack-unity
-
websocket-sharp
-
Copyright (c) 2010-2016 sta.blockhead
-
The MIT License (MIT)
-
Assets/Packages/websocket-sharp
License
- Apache License, Version 2.0
- Original Developer: (DWANGO ARTIFICIAL INTELLIGENCE LABORATORY)
Notice
If you created intelligent agents, please let me know about it to "masayoshi_nakamura@dwango.co.jp". We will make showcase varied intelligent agents.