Topics Covered

We are a reading group that explores the cutting-edge of Reinforement Learning and we summarize topics in RL literature, feel free to check out our slides. You are invited to join next sessions: [Notify Me], no prep needed except for RL fundamentals.

Neural Memory

Review of recent approaches to short-term memory in the Reinforcement Learning setting.

TBA
TBA

Curiosity-driven Exploration by Self-supervised Prediction

The extrinsic rewards to the agent are extremely sparse. The authors of this paper turn to curiosity as instrisic reward signal and introduce the Intrinsic Curiosity Module (ICM).


Contrastive Learning of Structured World Models

This recent paper by Thomas Kipf, Elise van der Pol and Max Welling highlights different cognitive architectures 'how to see' in the Reinforcement Learning setting. Their proposed C-SWM model is composed of a CNN-based objectextractor, an MLP-based object encoder, a GNN-based relational transition model, and an object-factorized contrastive loss.


The Options framework

In Reinforcement Learning, the agent learns to perform some actions in an environment, which is what we call the policy. Instead of having one policy, the agent could instead have a policy over options. Every option contains its own policy and termination function. Suggested by Sutton and others as early as 2000, what is the current state of options?


MuZero: Mastering Atari, Go, Chess and Shogi by Planning with a Learned Model

Take an AI that learns to play Go and other board games better than any human using only self-play, now make it work for ATARI games. That is exactly what the authors of MuZero have achieved, and in doing so they beat the state-of-the-art on many of the ATARI games. We covered their algorithm in more detail and compare it to AlphaZero.


Learning Plannable Representations with Causal InfoGAN

This work by two students at UC Berkeley learns to navigate the environment without any objective function. The proposed model-based RL links observations and states in an unsupervised way, using the Infogan.


Introduction to our reading group