-
Notifications
You must be signed in to change notification settings - Fork 0
/
Copy pathresearch.html
219 lines (203 loc) · 15.7 KB
/
research.html
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
<!DOCTYPE html>
<html lang="en">
<head>
<meta charset="UTF-8">
<meta name="viewport" content="width=device-width, initial-scale=1.0">
<title>AI4CO</title>
<link rel="icon" type="image/svg" href="images/ai4co_icon.svg">
<link rel="stylesheet" href="css/main.css">
<link href="https://fonts.googleapis.com/css2?family=Open+Sans:wght@100;200;300;400;500&display=swap" rel="stylesheet">
<link href="https://fonts.googleapis.com/css2?family=Nanum+Myeongjo:wght@100&display=swap" rel="stylesheet">
</head>
<body>
<nav class="navbar">
<div class="navbar-brand">
<a href="/">
<!-- Place your logo here -->
<img src="images/ai4co_logo_white.png" alt="AI4CO Logo">
</a>
</div>
<div class="navbar-icon">Menu</div>
<div class="navbar-menu">
<a href="research.html">Research <svg data-v-67d7128c="" fill="none" xmlns="http://www.w3.org/2000/svg" viewBox="0 0 16 16" class="a-icon--arrow-north-east400 a-icon--text a-icon--no-align top-[0.05em] relative f-ui-1 ml-2 -mr-4" data-new="" aria-hidden="true" style="width: 1em; height: 1em;"><polygon data-v-67d7128c="" fill="currentColor" points="5 4.31 5 5.69 9.33 5.69 2.51 12.51 3.49 13.49 10.31 6.67 10.31 11 11.69 11 11.69 4.31 5 4.31"></polygon></svg></a>
<a href="https://join.slack.com/t/rl4co/shared_invite/zt-1ytz2c1v4-0IkQ8NQH4TRXIX8PrRmDhQ">Community <svg data-v-67d7128c="" fill="none" xmlns="http://www.w3.org/2000/svg" viewBox="0 0 16 16" class="a-icon--arrow-north-east400 a-icon--text a-icon--no-align top-[0.05em] relative f-ui-1 ml-2 -mr-4" data-new="" aria-hidden="true" style="width: 1em; height: 1em;"><polygon data-v-67d7128c="" fill="currentColor" points="5 4.31 5 5.69 9.33 5.69 2.51 12.51 3.49 13.49 10.31 6.67 10.31 11 11.69 11 11.69 4.31 5 4.31"></polygon></svg></a>
<a href="blog.html">Blog <svg data-v-67d7128c="" fill="none" xmlns="http://www.w3.org/2000/svg" viewBox="0 0 16 16" class="a-icon--arrow-north-east400 a-icon--text a-icon--no-align top-[0.05em] relative f-ui-1 ml-2 -mr-4" data-new="" aria-hidden="true" style="width: 1em; height: 1em;"><polygon data-v-67d7128c="" fill="currentColor" points="5 4.31 5 5.69 9.33 5.69 2.51 12.51 3.49 13.49 10.31 6.67 10.31 11 11.69 11 11.69 4.31 5 4.31"></polygon></svg></a>
<a href="about.html">About <svg data-v-67d7128c="" fill="none" xmlns="http://www.w3.org/2000/svg" viewBox="0 0 16 16" class="a-icon--arrow-north-east400 a-icon--text a-icon--no-align top-[0.05em] relative f-ui-1 ml-2 -mr-4" data-new="" aria-hidden="true" style="width: 1em; height: 1em;"><polygon data-v-67d7128c="" fill="currentColor" points="5 4.31 5 5.69 9.33 5.69 2.51 12.51 3.49 13.49 10.31 6.67 10.31 11 11.69 11 11.69 4.31 5 4.31"></polygon></svg></a>
</div>
</nav>
<script src="js/nav.js"></script>
<section class="hero hero-nobg">
<div class="hero-content hero-content-nobg">
<p id="hero-minititle">Research</p>
<p>RL4CO: An extensive Reinforcement Learning (RL) for Combinatorial Optimization (CO) benchmark</p>
<a href="https://github.com/ai4co/rl4co" class="btn"><img src="images/github_logo.svg" alt="github_logo" width="14px" height="14px"> GitHub</a>
<a href="https://arxiv.org/abs/2306.17100" class="btn btn-nobd">View arXiv Preprint <svg data-v-67d7128c="" fill="none" xmlns="http://www.w3.org/2000/svg" viewBox="0 0 16 16" class="a-icon--arrow-north-east400 a-icon--text a-icon--no-align top-[0.05em] relative f-ui-1 ml-2 -mr-4" data-new="" aria-hidden="true" style="width: 1em; height: 1em;"><polygon data-v-67d7128c="" fill="currentColor" points="5 4.31 5 5.69 9.33 5.69 2.51 12.51 3.49 13.49 10.31 6.67 10.31 11 11.69 11 11.69 4.31 5 4.31"></polygon></svg></a>
</div>
</section>
<section class="updates-section">
<div class="container">
<div class="updates-grid">
<div class="update-item">
<div class="update-content">
<h3>You can easily install and try RL4CO form PyPI</h3>
<code>pip install rl4co</code><button class="copy-btn" onclick="copyToClipboard('#code')"><img src="images/copy-icon.svg" alt="Copy"></button>
</div>
</div>
<div class="update-item">
<div class="update-content">
<h3>Explor RL4CO's models, environments, and APIs</h3>
<a class="update-link" href="https://rl4co.readthedocs.io/en/latest/">Read the documents</a><svg data-v-67d7128c="" fill="none" xmlns="http://www.w3.org/2000/svg" viewBox="0 0 16 16" class="a-icon--arrow-north-east400 a-icon--text a-icon--no-align top-[0.05em] relative f-ui-1 ml-2 -mr-4" data-new="" aria-hidden="true" style="width: 1em; height: 1em;"><polygon data-v-67d7128c="" fill="currentColor" points="5 4.31 5 5.69 9.33 5.69 2.51 12.51 3.49 13.49 10.31 6.67 10.31 11 11.69 11 11.69 4.31 5 4.31"></polygon></svg>
</div>
</div>
<div class="update-item">
<div class="update-content">
<h3>Connect with peers to share insights on RL4CO</h3>
<a class="update-link" href="https://join.slack.com/t/rl4co/shared_invite/zt-1ytz2c1v4-0IkQ8NQH4TRXIX8PrRmDhQ">Join us with Slack</a><svg data-v-67d7128c="" fill="none" xmlns="http://www.w3.org/2000/svg" viewBox="0 0 16 16" class="a-icon--arrow-north-east400 a-icon--text a-icon--no-align top-[0.05em] relative f-ui-1 ml-2 -mr-4" data-new="" aria-hidden="true" style="width: 1em; height: 1em;"><polygon data-v-67d7128c="" fill="currentColor" points="5 4.31 5 5.69 9.33 5.69 2.51 12.51 3.49 13.49 10.31 6.67 10.31 11 11.69 11 11.69 4.31 5 4.31"></polygon></svg>
</div>
</div>
</div>
</div>
</section>
<hr>
<section class="updates-section updates-section-title">
<h2>RL4CO provides a unified framework for RL-based CO algorithms, and to facilitate reproducible research in this field, decoupling the science from the engineering.</h2>
<div class="container">
<div class="updates-grid">
<div class="update-item">
<p>RL4CO is built upon:</p>
</div>
<div class="update-item">
<p>TorchRL: official PyTorch framework for RL algorithms and vectorized environments on GPUs;</p>
<p>TensorDict: a library to easily handle heterogeneous data such as states, actions and rewards;</p>
<p>PyTorch Lightning: a lightweight PyTorch wrapper for high-performance AI research;</p>
<p>Hydra: a framework for elegantly configuring complex applications;</p>
</div>
</div>
</div>
</section>
<hr>
<section class="updates-section updates-section-title">
<h2>RL4CO aims to decouple the major components of the autoregressive policy of NCO and its training routine while prioritizing reusability.</h2>
<div class="container">
<div class="updates-grid">
<div class="update-item">
<p class="update-item-lefttext">
This module takes the problem and constructs solutions autoregressively. The policy consists of the following components:
<span class="codetext">Init Embedding</span>, <span class="codetext">Encoder</span>, <span class="codetext">Context Embedding</span>, and <span class="codetext">Decoder</span>. Each of these components is
designed as an independent module for easy integration. Our policy design is flexible enough to reimplement state-of-the-art
autoregressive policies, including AM, POMO, and Sym-NCO, for various CO problems such as TSP, CVRP, OP, PCTSP, PDP, and mTSP,
to name a few.
</p>
<p class="update-item-lefttext">
This module fully specifies the given problem and updates the problem construction steps based on the input action.
When implementing the <span class="codetext">environment</span>, we focus on parallel execution of rollouts (i.e., problem-solving) while
maintaining statelessness in updating every step of solution decoding. These features are essential for ensuring
the reproducibility of NCO and supporting "look-back" decoding schemes such as Monte-Carlo Tree Search. <br><br>
Our environment implementation is based on <span class="codetext">TorchRL</span>, an open-source RL
library for <span class="codetext">PyTorch</span>, which aims at high modularity and good runtime performance,
especially on GPUs. This design choice makes the <span class="codetext">Environment</span> implementation standalone, even outside of RL4CO,
and consistently empowered by a community-supporting library -- <span class="codetext">TorchRL</span>.
</p>
<p class="update-item-lefttext">
This module defines the routine that takes the <span class="codetext">Policy</span>, <span class="codetext">Environment</span>, and problem instances and generates
the gradients of the policy (and possibly the critic for actor-critic methods). We intentionally decouple the routines for
gradient computations and parameter updates to support modern training practices.
</p>
<p class="update-item-lefttext">
Training a single NCO model is typically computationally demanding, especially since most CO problems are NP-hard.
Therefore, implementing a modernized training routine becomes crucial. To this end, we implement the <span class="codetext">Trainer</span>
using <span class="codetext">Lightning</span>, which seamlessly supports features of modern training
pipelines, including logging, checkpoint management, automatic mixed-precision training, various hardware acceleration
supports (e.g., CPU, GPU, TPU, and Apple Silicon), multi-GPU support, and even multi-machine expansion. We have found
that using mixed-precision training significantly decreases training time without sacrificing NCO solver quality and
enables leveraging recent routines such as FlashAttention.
</p>
<p class="update-item-lefttext">
Optionally, but usefully, we adopt <span class="codetext">Hydra</span>, an open-source Python framework that enables
hierarchical config management. It promotes modularity, scalability, and reproducibility, making it easier to manage
complex configurations and experiments with different settings and maintain consistency across different environments.
An overview of RL4CO code implementation is visualized in the overview figure.
</p>
</div>
</div>
</div>
</section>
<hr>
<section class="updates-section updates-section-title">
<h2>Contribute</h2>
<div class="container">
<div class="updates-grid">
<div class="update-item">
<p class="update-item-lefttext">Have a suggestion, request, or found a bug? Feel free to open an issue or submit a pull request. We welcome and look forward to all contributions to RL4CO! We are also on Slack if you have any questions or would like to discuss RL4CO with us. We are open to collaborations and would love to hear from you!</p>
</div>
</div>
</div>
</section>
<hr>
<section class="updates-section updates-section-title">
<h2>Cite Us</h2>
<div class="container">
<div class="updates-grid">
<div class="update-item">
<p class="update-item-lefttext">If you find RL4CO valuable for your research or applied projects:</p>
<pre>
@article{berto2023rl4co,
title = {{RL4CO}: an Extensive Reinforcement Learning for Combinatorial Optimization Benchmark},
author={Federico Berto and Chuanbo Hua and Junyoung Park and Minsu Kim and Hyeonah Kim and Jiwoo Son and Haeyeon Kim and Joungho Kim and Jinkyoo Park},
journal={arXiv preprint arXiv:2306.17100},
year={2023},
url = {https://github.com/ai4co/rl4co}
</pre>
</div>
</div>
</div>
</section>
</body>
<footer class="site-footer">
<div class="container">
<div class="footer-logo">
<img src="images/ai4co_logo_white.png" alt="AI4CO Logo"/>
</div>
<div class="footer-columns">
<div class="footer-column">
<h4>Research</h4>
<ul>
<li>
<a href="research.html">RL4CO<svg data-v-67d7128c="" fill="none" xmlns="http://www.w3.org/2000/svg" viewBox="0 0 16 16" class="a-icon--arrow-north-east400 a-icon--text a-icon--no-align top-[0.05em] relative f-ui-1 ml-2 -mr-4" data-new="" aria-hidden="true" style="width: 1em; height: 1em;"><polygon data-v-67d7128c="" fill="currentColor" points="5 4.31 5 5.69 9.33 5.69 2.51 12.51 3.49 13.49 10.31 6.67 10.31 11 11.69 11 11.69 4.31 5 4.31"></polygon></svg></a>
</li>
</ul>
</div>
<div class="footer-column">
<h4>Contribution</h4>
<ul>
<li><a href="https://github.com/ai4co">GitHub<svg data-v-67d7128c="" fill="none" xmlns="http://www.w3.org/2000/svg" viewBox="0 0 16 16" class="a-icon--arrow-north-east400 a-icon--text a-icon--no-align top-[0.05em] relative f-ui-1 ml-2 -mr-4" data-new="" aria-hidden="true" style="width: 1em; height: 1em;"><polygon data-v-67d7128c="" fill="currentColor" points="5 4.31 5 5.69 9.33 5.69 2.51 12.51 3.49 13.49 10.31 6.67 10.31 11 11.69 11 11.69 4.31 5 4.31"></polygon></svg></a></li>
<li><a href="https://join.slack.com/t/rl4co/shared_invite/zt-1ytz2c1v4-0IkQ8NQH4TRXIX8PrRmDhQ">Slack<svg data-v-67d7128c="" fill="none" xmlns="http://www.w3.org/2000/svg" viewBox="0 0 16 16" class="a-icon--arrow-north-east400 a-icon--text a-icon--no-align top-[0.05em] relative f-ui-1 ml-2 -mr-4" data-new="" aria-hidden="true" style="width: 1em; height: 1em;"><polygon data-v-67d7128c="" fill="currentColor" points="5 4.31 5 5.69 9.33 5.69 2.51 12.51 3.49 13.49 10.31 6.67 10.31 11 11.69 11 11.69 4.31 5 4.31"></polygon></svg></a></li>
</ul>
</div>
<!-- <div class="footer-column">
<h4>About</h4>
<ul>
<li><a href="#">Target</a></li>
<li><a href="#">Developer</a></li>
<li><a href="#">Cooperators</a></li>
</ul>
</div> -->
</div>
</div>
<div class="container">
<div class="footer-bottom">
<div class="footer-info">
<p>AI4CO © 2023</p>
<!-- <ul class="legal-links">
<li><a href="#">Privacy policy</a></li>
<li><a href="#">License</a></li>
</ul> -->
</div>
<div class="footer-social">
<a href="https://github.com/ai4co">GitHub</a>
<a href="https://join.slack.com/t/rl4co/shared_invite/zt-1ytz2c1v4-0IkQ8NQH4TRXIX8PrRmDhQ">Slack</a>
</div>
</div>
</div>
</footer>
</html>