Atari Alien Environment

Overview

In this game based, loosely, on the movie of the same name, you have to move through a maze (the halls of your ship in the manual), ala Pac-Man, collecting dots (destroying alien eggs).

If you collect the power dot (pulsar), you can kill any of the three aliens, for a short time. There are only three enemies in the maze at a time, there is a bonus item at times and only one power dot (pulsar) at a time. When you grab the pulsar, it will next appear in one of two other spots.

After you clear one level, you get a bonus game. You have to move up the screen to the prize at the top past several aliens, reminiscent of Freeway. You do not lose a man if you fail but you only have eight seconds to do it then you are off the the next, harder level.

Description from RetroGames

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
6371.3 Human Massively Parallel Methods for Deep Reinforcement Learning
6022.9 Rainbow Rainbow: Combining Improvements in Deep Reinforcement Learning
1997.5 Distributional DQN Rainbow: Combining Improvements in Deep Reinforcement Learning
1486.5 DuDQN Dueling Network Architectures for Deep Reinforcement Learning
1334.7 Prioritized DDQN (rank, tuned) Prioritized Experience Replay
1191.0 Prioritized DQN (rank) Prioritized Experience Replay
1033.4 DDQN (tuned) Deep Reinforcement Learning with Double Q-learning
945.3 A3C LSTM Asynchronous Methods for Deep Reinforcement Learning
900.5 Prioritized DDQN (prop, tuned) Prioritized Experience Replay
823.7 PDD DQN Dueling Network Architectures for Deep Reinforcement Learning
813.54 Gorila DQN Massively Parallel Methods for Deep Reinforcement Learning
621.6 DDQN Deep Reinforcement Learning with Double Q-learning
570.2 DQN Massively Parallel Methods for Deep Reinforcement Learning
518.4 A3C FF Asynchronous Methods for Deep Reinforcement Learning
182.1 A3C FF 1 day Asynchronous Methods for Deep Reinforcement Learning
128.3 Random Massively Parallel Methods for Deep Reinforcement Learning

No-op Starts

Result Algorithm Source
15962.1 IMPALA (deep) IMPALA: Scalable Distributed Deep-RL with Importance Weighted Actor-Learner Architectures
12689.1 Reactor The Reactor: A fast and sample-efficient Actor-Critic agent for Reinforcement Learning
9983 QR-DQN-0 Distributional Reinforcement Learning with Quantile Regression
9491.7 Rainbow Rainbow: Combining Improvements in Deep Reinforcement Learning
7127.7 Human Dueling Network Architectures for Deep Reinforcement Learning
7022 IQN Implicit Quantile Networks for Distributional Reinforcement Learning
6875.4 Human Human-level control through deep reinforcement learning
6482.1 Reactor The Reactor: A fast and sample-efficient Actor-Critic agent for Reinforcement Learning
6163 DuDQN Noisy Networks for Exploration
5778 NoisyNet DuDQN Noisy Networks for Exploration
4871 QR-DQN-1 Distributional Reinforcement Learning with Quantile Regression
4461.4 DuDQN Dueling Network Architectures for Deep Reinforcement Learning
4199.4 Reactor ND The Reactor: A fast and sample-efficient Actor-Critic agent for Reinforcement Learning
4055.8 Distributional DQN Rainbow: Combining Improvements in Deep Reinforcement Learning
3941.0 PDD DQN Dueling Network Architectures for Deep Reinforcement Learning
3747.7 DDQN A Distributional Perspective on Reinforcement Learning
3197.1 ACKTR Scalable trust-region method for deep reinforcement learning using Kronecker-factored approximation
3166 C51 A Distributional Perspective on Reinforcement Learning
3069 DQN Human-level control through deep reinforcement learning
2907.3 DDQN Deep Reinforcement Learning with Double Q-learning
2620.53 Gorila DQN Massively Parallel Methods for Deep Reinforcement Learning
2404 DQN Noisy Networks for Exploration
2403 NoisyNet DQN Noisy Networks for Exploration
2344.6 IMPALA (deep, multitask) IMPALA: Scalable Distributed Deep-RL with Importance Weighted Actor-Learner Architectures
2027 A3C Noisy Networks for Exploration
1899 NoisyNet A3C Noisy Networks for Exploration
1620.0 DQN A Distributional Perspective on Reinforcement Learning
1536.05 IMPALA (shallow) IMPALA: Scalable Distributed Deep-RL with Importance Weighted Actor-Learner Architectures
939.2 Linear Human-level control through deep reinforcement learning
227.8 Random Human-level control through deep reinforcement learning
103.2 Contingency Human-level control through deep reinforcement learning

Normal Starts

Result Algorithm Source
1850.3 PPO Proximal Policy Optimization Algorithm
1655.4 ACER Proximal Policy Optimization Algorithm
1141.7 A2C Proximal Policy Optimization Algorithm