-
Notifications
You must be signed in to change notification settings - Fork 1
/
Copy pathSL_2022_projects.html
237 lines (194 loc) · 11.2 KB
/
SL_2022_projects.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
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
<!DOCTYPE HTML>
<!--
Read Only by HTML5 UP
html5up.net | @n33co
Free for personal and commercial use under the CCA 3.0 license (html5up.net/license)
-->
<html>
<head>
<title>Sequential Learning</title>
<meta charset="utf-8" />
<meta name="viewport" content="width=device-width, initial-scale=1" />
<!--[if lte IE 8]><script src="assets/js/ie/html5shiv.js"></script><![endif]-->
<link rel="stylesheet" href="assets/css/main.css" />
<!--[if lte IE 8]><link rel="stylesheet" href="assets/css/ie8.css" /><![endif]-->
</head>
<body>
<!-- Header -->
<section id="header">
<header>
<span class="image avatar"><img src="images/avatar2.jpg" alt="" /></span>
<h1 id="logo"><a href="https://remydegenne.github.io/SL_2022">Sequential Learning (2022)</a></h1>
<p>Rémy Degenne</p>
</header>
<footer>
<ul class="icons">
<li><a href="#" class="icon fa-github"><span class="label">Github</span></a></li>
</ul>
</footer>
</section>
<!-- Wrapper -->
<div id="wrapper">
<!-- Main -->
<div id="main">
<!-- Three-->
<section id="one">
<div class="container">
<h3>Sequential learning projects</h3>
<p>
The project consists in reading, understanding and explaining a research paper about a topic related to the class [you don’t need to understand all the proofs when there are some, but you do need to understand the results]. If possible, you are encouraged to implement the algorithm described yourself and report your own findings on close or different benchmarks than those presented in the paper.
</p>
<p>
What is expected? A short report (8 pages) explaining the topic in your own words (no copy-paste from the paper your study!), relating it to the course, discussing links with other algorithms you know, etc., reporting the findings of your own experiments,our opinion about the paper (possible weaknesses ?).
In short any personal thoughts inspired by the paper.
</p>
<p>
You can team up for the project, groups of two students are welcome. 1 group / project maximum.
</p>
<p>
Possible format: you may use the NeurIPS template if you are familiar with latex, but any other format is accepted. Using Overleaf allows you to write latex in a navigator with no installation on your computer, and the NeurIPS template is already there:
<a href="https://fr.overleaf.com/latex/templates/neurips-2022/kxymzbjpwsqx">https://fr.overleaf.com/latex/templates/neurips-2022/kxymzbjpwsqx</a>
</p>
<p>
<b>Deadline: send your report to me by email by January 26nd, 2023 at 11:59pm. </b>
You can either submit a PDF file with your report or a zip file with your PDF report and source code to run your experiments (a github link in the report is also OK). In either case, the name of your submission should be SL22_Name1_Name2.pdf
</p>
<h3>List of possible projects</h3>
You can also propose your own, please email me for approval with a short description of the project.
<h4>Bandits</h4>
<p>
<b>Bootstrapping and bandits</b>: researchers have developed alternatives to UCB or Thompson Sampling with an exploration that is more data-driven and directly uses the history of observed reward. This paper reveals that standard statistical bootstrap method cannot work, and propose a fix
</br><a href="https://arxiv.org/abs/1811.05154">https://arxiv.org/abs/1811.05154 </a>
</br>Project chosen by: Alexandre Dezalot, Clément Jurat-Pe tiadou
</p>
<p>
<b>Bandits with delayed rewards</b>: In real-world problems, the reward is often observed after some random delay. This paper presents a way to cope with a known delay distribution:
<a href="https://arxiv.org/abs/1706.09186">https://arxiv.org/abs/1706.09186</a>. Maybe you want to try your own Thompson Sampling variant?
</br>Project chosen by: Caroline Charreteur, Jean-Paul Duboux
</p>
<p>
<b>Non-stationary bandits</b>: Assuming the rewards are i.i.d. is often a big limitation in practice, and researchers have studied the so-called piecewise i.i.d. models, in which there can be some breakpoints in which the rewards distributions change abruptly. This paper proposes two ways to cope with non-stationarity, that you can understand and compare :
</br><a href="https://arxiv.org/abs/0805.3415">https://arxiv.org/abs/0805.3415 </a>
</br>Project chosen by: Zakaria Farah, Ilyas Ouardi
</p>
<p>
<b>Thompson Sampling for logistic bandits</b>: a logistic model is particularly well suited to model binary rewards and take features into account and could be useful for contextual recommendation. This paper discusses the implementation of Thompson Sampling in this setting:
</br><a href="http://papers.neurips.cc/paper/7713-pg-ts-improved-thompson-sampling-for-logistic-contextual-bandits">http://papers.neurips.cc/paper/7713-pg-ts-improved-thompson-sampling-for-logistic-contextual-bandits</a>
</br>Project chosen by:
</p>
<!--
<p>
<b>Bandits without rewards</b>: Is it possible to minimize regret in a bandit model and simultaneously find the best arm with high probability?
</br><a href="https://arxiv.org/abs/1810.04088">https://arxiv.org/abs/1810.04088 </a>
</br>Project chosen by:
</p>
-->
<h4>RL algorithms, miscellaneous</h4>
<p>
<b>Conservative policy iteration</b>: How can we do policy iteration if the policies cannot be evaluated exactly?
</br><a href="https://people.eecs.berkeley.edu/~pabbeel/cs287-fa09/readings/KakadeLangford-icml2002.pdf">https://people.eecs.berkeley.edu/~pabbeel/cs287-fa09/readings/KakadeLangford-icml2002.pdf</a>
</br>Project chosen by: Victor Cilleros, Victor Niaussat
</p>
<p>
<b>Kernel-Based Reinforcement Learning</b>: A stable method of solving MDPs with continuous states when a simulator is available, with asymptotic convergence guarantees.
</br><a href="https://link.springer.com/article/10.1023/A:1017928328829">https://link.springer.com/article/10.1023/A:1017928328829</a>
</br>Project chosen by: Alexandre Antunes, Mounir Guerrab
</p>
<p>
<b>Transfer in Reinforcement Learning</b>: how to use reinforcement learning to learn several tasks at the same time, and get a more adaptable algorithm?
</br><a href="https://proceedings.neurips.cc/paper/2017/file/350db081a661525235354dd3e19b8c05-Paper.pdf">https://proceedings.neurips.cc/paper/2017/file/350db081a661525235354dd3e19b8c05-Paper.pdf</a>
</br>Project chosen by: Pierre Mauron, Victor Calamy
</p>
<h4>Improving the DQN algorithm</h4>
<p>
<b>Deep Reinforcement Learning with Double Q-learning</b>: A way to improve Deep Q Learning.
</br><a href="https://arxiv.org/pdf/1509.06461.pdf">https://arxiv.org/pdf/1509.06461.pdf</a>
</br>Project chosen by: Mohamed Ben Ali, Mohamed Makni
</p>
<p>
<b>Prioritized Experience Replay</b>: Another way to improve Deep Q Learning.
</br><a href="https://arxiv.org/abs/1511.05952">https://arxiv.org/abs/1511.05952</a>
</br>Project chosen by:
</p>
<p>
<b>Dueling Network Architectures for Deep Reinforcement Learning</b>: Again another way to improve Deep Q Learning.
</br><a href="https://arxiv.org/pdf/1511.06581.pdf">https://arxiv.org/pdf/1511.06581.pdf</a>
</br>Project chosen by:
</p>
<h4>Policy Optimization / Policy Gradient </h4>
<p>
<b>Policy optimization via the cross-entropy method</b>: The cross-entropy (CE) method is a Monte Carlo method for importance sampling and optimization, that can be used to optimize the value function in reinforcement learning.
</br><a href="https://www.aaai.org/Papers/ICML/2003/ICML03-068.pdf">https://www.aaai.org/Papers/ICML/2003/ICML03-068.pdf</a>
</br>Project chosen by: Jeremy Jean
</p>
<!--
<p>
<b>Bandits and Policy gradients algorithms</b>: When bandits meets Policy gradient for Optimistic Policy Optimization :
</br><a href="https://t3p.github.io/icml19/">https://t3p.github.io/icml19/</a>
</br>Project chosen by:
</p>
-->
<p>
<b>A Natural Policy Gradient</b>: Policy gradient methods perform gradient ascent on the value function to optimize the policy. The natural policy gradient allows us to improve the convergence speed of this approach.
</br><a href="https://papers.nips.cc/paper/2001/file/4b86abe48d358ecf194c56c69108433e-Paper.pdf">https://papers.nips.cc/paper/2001/file/4b86abe48d358ecf194c56c69108433e-Paper.pdf</a>
</br>Project chosen by:
</p>
<h4>Exploration in RL </h4>
<p>
<b>Posterior Sampling for RL</b>: the Thompson Sampling principle can be naturally extended to RL, provided one is able to maintain a posterior on the parameters of a Markov Decision Process. You can try to understand how it’s done here, and what theoretical guarantees the authors are able to give for this new exploration strategy.
</br><a href="https://arxiv.org/abs/1306.0940">https://arxiv.org/abs/1306.0940 </a>
</br>Project chosen by: Céline Wang
</p>
<p>
<b>Bootstrap DQN</b>: the DQN algorithm is often used in conjunction with an epsilon-greedy policy to perform exploration. The Bootstrap DQN alternative introduces some alternative, randomized exploration :
</br><a href="https://arxiv.org/abs/1602.04621">https://arxiv.org/abs/1602.04621 </a>
</br>Project chosen by:
</p>
<p>
<b>Exploration: A Study of Count-Based Exploration for Deep Reinforcement Learning</b>: Improving exploration in Deep RL using bonuses built with hash codes of the states.
</br><a href="https://arxiv.org/pdf/1611.04717.pdf">https://arxiv.org/pdf/1611.04717.pdf </a>
</br>Project chosen by:
</p>
<p>
<b>Unifying Count-Based Exploration and Intrinsic Motivation</b>: Improving exploration in RL using bonuses built with pseudo-counts based on a density model.
</br><a href="https://arxiv.org/pdf/1606.01868.pdf">https://arxiv.org/pdf/1606.01868.pdf </a>
</br>Project chosen by:
</p>
<p>
<b>AlphaZero</b>:
</br><a href="https://arxiv.org/abs/1712.01815">https://arxiv.org/abs/1712.01815</a>, <a href="http://www.dcsc.tudelft.nl/~sc4081/2018/assign/pap/alphago_paper1.pdf">http://www.dcsc.tudelft.nl/~sc4081/2018/assign/pap/alphago_paper1.pdf</a>
</br>Project chosen by: Kenza Makhlouf
</p>
<h4>Miscellaneous </h4>
<p>
<b>From TD-Gammon to DQN</b> : a journey in the history of RL. The goal of this project is to understand the algorithm that lead to the first big success of RL and already used neural networks, and contrast it with the DQN algorithm.
</br><a href="https://bkgm.com/articles/tesauro/TDGammonAchievesMasterLevelPlay.pdf">https://bkgm.com/articles/tesauro/TDGammonAchievesMasterLevelPlay.pdf </a>
</br>Project chosen by: Judicael Leger, Alexandre Caumette
</p>
<p>
<b>Bandits to manage inventory problems</b>: How bandit models can be useful to solve inventory management problems when the demand is unknown.
</br><a href="https://arxiv.org/abs/1905.04337">https://arxiv.org/abs/1905.04337 </a>
</br>Project chosen by: Amine Belblidia, Julien Airault
</p>
</div>
</section>
</div>
<!-- Footer -->
<section id="footer">
<div class="container">
<ul class="copyright">
<li>© Untitled. All rights reserved.</li><li>Design: <a href="http://html5up.net">HTML5 UP</a></li>
</ul>
</div>
</section>
</div>
<!-- Scripts -->
<script src="assets/js/jquery.min.js"></script>
<script src="assets/js/jquery.scrollzer.min.js"></script>
<script src="assets/js/jquery.scrolly.min.js"></script>
<script src="assets/js/skel.min.js"></script>
<script src="assets/js/util.js"></script>
<!--[if lte IE 8]><script src="assets/js/ie/respond.min.js"></script><![endif]-->
<script src="assets/js/main.js"></script>
</body>
</html>