ps:简单记录ICML2018论文研讨会内容
2018.7.23
Competitive Multi-agent Inverse Reinforcement Learning with Sub-optimal Demonstrations. http://proceedings.mlr.press/v80/wang18d/wang18d.pdf
- 零和博弈(GAN受此启发)和逆强化学习
Learning to Explore via Meta-Policy Gradient. http://proceedings.mlr.press/v80/xu18d/xu18d.pdf
- 元策略梯度
QMIX: Monotonic Value Function Factorisation for Deep Multi-Agent Reinforcement Learning. http://proceedings.mlr.press/v80/rashid18a/rashid18a.pdf
- 值函数分解
2018.7.24
Ray: A Distributed Framework for Emerging AI Applications. Arxiv. https://arxiv.org/pdf/1712.05889.pdf
- 伯克利分布式工具,分享者并没有讲清楚如何部署分布式
Katyusha X: Practical Momentum Method for Stochastic Sum-of-Nonconvex Optimization. ICML 2018. http://proceedings.mlr.press/v80/allen-zhu18a/allen-zhu18a.pdf
- 非凸优化
Self-Imitation Learning. ICML 2018. http://proceedings.mlr.press/v80/oh18b/oh18b.pdf
- 自模仿学习
2018.7.27
Mix & Match - Agent Curricula for Reinforcement learning. ICML 2018. http://proceedings.mlr.press/v80/czarnecki18a/czarnecki18a.pdf
- transfer learning用于强化学习
- k越大,模型吸收前面模型的内容越多,训练复杂度越高
Self-Consistent Trajectory Autoencoder: Hierarchical Reinforcement Learning with Trajectory Embeddings. ICML 2018. http://proceedings.mlr.press/v80/co-reyes18a/co-reyes18a.pdf
- 类似于VAE用于分层强化学习
State Abstractions for Lifelong Reinforcement Learning. ICML 2018. http://proceedings.mlr.press/v80/abel18a/abel18a.pdf
- 终身强化学习 相当于任务可迁移
2018.7.28
Efficient Neural Architecture Search via Parameter Sharing. ICML 2018. http://proceedings.mlr.press/v80/pham18a/pham18a.pdf
- 在NAS基础上做改进,对于给定的神经网络模块,建立DAG图,具体算法有待继续研究
- Google Brain的insight很好,但是还很weak
Implicit Quantile Networks for Distributional Reinforcement Learning. ICML 2018. http://proceedings.mlr.press/v80/dabney18a/dabney18a.pdf
- 本文分成Quantile和Distributional,可以看下作者之前两篇工作
2018.7.29
Bayesian Optimization of Combinatorial Structures. ICML 2018. http://proceedings.mlr.press/v80/baptista18a/baptista18a.pdf
- 没听懂,只是取得部分进展
Visualizing and Understanding Atari Agents. ICML 2018. http://proceedings.mlr.press/v80/greydanus18a/greydanus18a.pdf
- 高斯模糊某一片,看看这块区域对于Q值的影响
Policy Optimization with Demonstrations. ICML 2018. http://proceedings.mlr.press/v80/kang18a/kang18a.pdf
- 没怎么听
2018.7.30
Fully Decentralized Multi-Agent Reinforcement Learning with Networked Agents. ICML 2018. http://proceedings.mlr.press/v80/zhang18n/zhang18n.pdf
- 智能体之间的通信,随机选择子图,理论早已经弄好,然后实验简单设计
Structured Evolution with Compact Architectures for Scalable Policy Optimization. ICML 2018. http://proceedings.mlr.press/v80/choromanski18a/choromanski18a.pdf
- Google brain的,讲了一堆矩阵概念,理论解释不清楚,实验完备
Using Reward Machines for High-Level Task Specification and Decomposition in Reinforcement Learning. ICML 2018. http://proceedings.mlr.press/v80/icarte18a/icarte18a.pdf
- 用状态机完成与环境交互一次,就能完成多任务的reward计算
2018.7.31
Essentially No Barriers in Neural Network Energy Landscape. ICML 2018.
http://proceedings.mlr.press/v80/draxler18a/draxler18a.pdf- 局部最优点连线
Time Limits in Reinforcement Learning. ICML 2018. http://proceedings.mlr.press/v80/pardo18a/pardo18a.pdf
- 考虑有限步长
Obfuscated Gradients Give a False Sense of Security: Circumventing Defenses to Adversarial Examples. ICML 2018. http://proceedings.mlr.press/v80/athalye18a/athalye18a.pdf
- best paper对抗的7篇中未被攻克的
2018.8.1
Learning with Abandonment. ICML 2018.
http://proceedings.mlr.press/v80/schmit18a/schmit18a.pdf- 在推荐系统中用强化学习,设计了一个用户容忍度theta
Latent Space Policies for Hierarchical Reinforcement Learning. ICML 2018. http://proceedings.mlr.press/v80/haarnoja18a/haarnoja18a.pdf
- 分层强化学习主要是解决解决系数reward或者复杂情况
- 这篇文章文不对标题的分层强化学习
Coordinated Exploration in Concurrent Reinforcement Learning. ICML 2018. http://proceedings.mlr.press/v80/dimakopoulou18a/dimakopoulou18a.pdf
- 提出了seed算法,对比了之前的UCB和辛普森采样,没有解释清楚Concurrent多智能体协同运作
2018.8.2
Clipped Action Policy Gradient. ICML 2018. http://proceedings.mlr.press/v80/fujita18a/fujita18a.pdf
- 求策略梯度的时候用alpha和beta截断,是无偏估计
An Inference-Based Policy Gradient Method for Learning Options. ICML 2018. http://proceedings.mlr.press/v80/smith18a/smith18a.pdf
- 分层强化学习领域的一篇文章与
- 与ICML2017的A Laplacian Framework for Option Discovery
in Reinforcement Learning算法类似,实验也有比较
2018.8.3
Universal Planning Networks: Learning Generalizable Representations for Visuomotor Control. ICML 2018. http://proceedings.mlr.press/v80/srinivas18b/srinivas18b.pdf
- 引出对state抽象,做一个model-based,model-based与model-free结合
Investigating Human Priors for Playing Video Games. ICML 2018. http://proceedings.mlr.press/v80/dubey18a/dubey18a.pdf
2018.8.4
Obfuscated Gradients Give a False Sense of Security: Circumventing Defenses to Adversarial Examples. ICML 2018. http://proceedings.mlr.press/v80/athalye18a/athalye18a.pdf
- ICML2018 best paper
- 原来的ICLR的基于梯度的防御机制主要有三种,分别是梯度破碎,随机梯度,多轮之后爆炸和消失梯度,一三的对抗方法是找一个不定点可导函数,第二个对抗方法是期望最大化
Addressing Function Approximation Error in Actor-Critic Methods. ICML 2018. http://proceedings.mlr.press/v80/fujimoto18a/fujimoto18a.pdf
- 状态抽象,类似于vae,对状态抽象再还原,最后再最小化作者提出的loss
Efficient Model-Based Deep Reinforcement Learning with Variational State Tabulation. ICML 2018. http://proceedings.mlr.press/v80/corneil18a/corneil18a.pdf
To be continue。。。。。