University Seminar Paper, not reviewed.
Survey/overview over the evolution of actor-critic reinforcement learning algorithms over time, including state-of-the-art of 2021 (time of writing the paper&poster).
Actor-critic algorithms are at the forefront of the fast evolving research field of reinforcement learning. Recent breakthroughs, like beating the world-champion in the game of Go in 2015, as well as reaching Grandmaster level in StarCraftII, used variants of the actor-critic framework. This catalysed the development of many variants and improvements, cross-influenced by the progress made on different model-free reinforcement learning algorithms. This work gives an overview over the state of the art of actor-critic algorithms, with a focus on popular benchmarks like the Arcade Learning Environment and MuJoCo.
First, an introduction to the actor-critic framework is given. Two popular benchmarks are described, followed by the history of different improvements and algorithms over the past 6 years. Finally, current state-of-the-art actor-critic methods are presented. The work concludes with a discussion about the difficulty of comparing different reinforcement learning algorithms.
This repository is linked a QR-Code on the poster, to provide the full seminar-paper to curious readers, as well as the included full list of references. The PDF version of the poster can be found here aswell.
Links: