Atari Surround Environment

Overview

Like its predecessor Blockade, the object of Surround is to maneuver a square across the screen, leaving a trail behind. A player wins by forcing the other player to crash into one of the trails. Twelve game variations include options allow for speed-up, diagonal movement, wrap-around, and “erase” (the choice to not draw at a given moment). In addition, the sprites can be set to operate at a slower speed, or progressively speed up through five speeds.

Description from Wikipedia

Performances of RL Agents

We list various reinforcement learning algorithms that were tested in this environment. These results are from RL Database. If this page was helpful, please consider giving a star!

Star

Human Starts

Result Algorithm Source
7.0 Rainbow Rainbow: Combining Improvements in Deep Reinforcement Learning
5.9 Prioritized DDQN (rank, tuned) Prioritized Experience Replay
5.4 Human Deep Reinforcement Learning with Double Q-learning
5.4 Human Dueling Network Architectures for Deep Reinforcement Learning
4.5 Distributional DQN Rainbow: Combining Improvements in Deep Reinforcement Learning
4.0 DuDQN Dueling Network Architectures for Deep Reinforcement Learning
1.9 DDQN (tuned) Deep Reinforcement Learning with Double Q-learning
-0.2 PDD DQN Dueling Network Architectures for Deep Reinforcement Learning
-0.9 Prioritized DDQN (prop, tuned) Prioritized Experience Replay
-5.3 Prioritized DQN (rank) Prioritized Experience Replay
-6.0 DQN Rainbow: Combining Improvements in Deep Reinforcement Learning
-7.6 DDQN Deep Reinforcement Learning with Double Q-learning
-8.3 A3C LSTM Asynchronous Methods for Deep Reinforcement Learning
-9.6 A3C FF 1 day Asynchronous Methods for Deep Reinforcement Learning
-9.7 A3C FF Asynchronous Methods for Deep Reinforcement Learning
-9.7 Random Deep Reinforcement Learning with Double Q-learning

No-op Starts

Result Algorithm Source
10 NoisyNet DuDQN Noisy Networks for Exploration
9.7 QR-DQN-0 Distributional Reinforcement Learning with Quantile Regression
9.7 Rainbow Rainbow: Combining Improvements in Deep Reinforcement Learning
9.4 IQN Implicit Quantile Networks for Distributional Reinforcement Learning
8.2 QR-DQN-1 Distributional Reinforcement Learning with Quantile Regression
7.56 IMPALA (deep) IMPALA: Scalable Distributed Deep-RL with Importance Weighted Actor-Learner Architectures
7.0 Reactor The Reactor: A fast and sample-efficient Actor-Critic agent for Reinforcement Learning
6.8 C51 A Distributional Perspective on Reinforcement Learning
6.7 Reactor The Reactor: A fast and sample-efficient Actor-Critic agent for Reinforcement Learning
6.5 Human Dueling Network Architectures for Deep Reinforcement Learning
6.2 Distributional DQN Rainbow: Combining Improvements in Deep Reinforcement Learning
4.4 DDQN A Distributional Perspective on Reinforcement Learning
4.4 DuDQN Dueling Network Architectures for Deep Reinforcement Learning
1.2 PDD DQN Dueling Network Architectures for Deep Reinforcement Learning
1 NoisyNet A3C Noisy Networks for Exploration
1 DuDQN Noisy Networks for Exploration
0.9 Reactor ND The Reactor: A fast and sample-efficient Actor-Critic agent for Reinforcement Learning
-1 NoisyNet DQN Noisy Networks for Exploration
-5.6 DQN A Distributional Perspective on Reinforcement Learning
-5.6 DQN Rainbow: Combining Improvements in Deep Reinforcement Learning
-6 DQN Noisy Networks for Exploration
-8 A3C Noisy Networks for Exploration
-8.13 IMPALA (shallow) IMPALA: Scalable Distributed Deep-RL with Importance Weighted Actor-Learner Architectures
-8.51 IMPALA (deep, multitask) IMPALA: Scalable Distributed Deep-RL with Importance Weighted Actor-Learner Architectures
-10.0 Random Dueling Network Architectures for Deep Reinforcement Learning

Normal Starts

| Result | Algorithm | Source | |——–|———–|——–|