Atari Defender Environment

Overview

Defender is a two-dimensional side-scrolling shooting game set on the surface of an unnamed planet. The player controls a space ship as it navigates the terrain, flying either to the left or right. A joystick controls the ship’s elevation, and five buttons control its horizontal direction and weapons. The object is to destroy alien invaders, while protecting astronauts on the landscape from abduction. Humans that are abducted return as mutants that attack the ship. Defeating the aliens allows the player to progress to the next level. Failing to protect the astronauts, however, causes the planet to explode and the level to become populated with mutants. Surviving the waves of mutants results in the restoration of the planet. Players are allotted three ships to progress through the game and are able to earn more by reaching certain scoring benchmarks. A ship is lost if it is hit by an enemy, or hit by an enemy projectile, or if a hyperspace jump goes wrong (as they randomly do). After exhausting all ships, the game ends.

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
233021.5 A3C LSTM Asynchronous Methods for Deep Reinforcement Learning
56533.0 A3C FF Asynchronous Methods for Deep Reinforcement Learning
47671.3 Rainbow Rainbow: Combining Improvements in Deep Reinforcement Learning
36242.5 A3C FF 1 day Asynchronous Methods for Deep Reinforcement Learning
34415.0 PDD DQN Dueling Network Architectures for Deep Reinforcement Learning
33996.0 DuDQN Dueling Network Architectures for Deep Reinforcement Learning
32246.0 Distributional DQN Rainbow: Combining Improvements in Deep Reinforcement Learning
27510.0 DDQN (tuned) Deep Reinforcement Learning with Double Q-learning
23666.5 Prioritized DDQN (rank, tuned) Prioritized Experience Replay
21093.5 Prioritized DDQN (prop, tuned) Prioritized Experience Replay
20634.0 Prioritized DQN (rank) Prioritized Experience Replay
15917.5 DQN Rainbow: Combining Improvements in Deep Reinforcement Learning
14296.0 Human Deep Reinforcement Learning with Double Q-learning
14296.0 Human Dueling Network Architectures for Deep Reinforcement Learning
8531.0 DDQN Deep Reinforcement Learning with Double Q-learning
1965.5 Random Deep Reinforcement Learning with Double Q-learning

No-op Starts

Result Algorithm Source
223025.0 Reactor The Reactor: A fast and sample-efficient Actor-Critic agent for Reinforcement Learning
185203.0 IMPALA (deep) IMPALA: Scalable Distributed Deep-RL with Importance Weighted Actor-Learner Architectures
181074.3 Reactor ND The Reactor: A fast and sample-efficient Actor-Critic agent for Reinforcement Learning
113128 Reactor The Reactor: A fast and sample-efficient Actor-Critic agent for Reinforcement Learning
58718.25 IMPALA (shallow) IMPALA: Scalable Distributed Deep-RL with Importance Weighted Actor-Learner Architectures
55492 NoisyNet A3C Noisy Networks for Exploration
55105.0 Rainbow Rainbow: Combining Improvements in Deep Reinforcement Learning
53537 IQN Implicit Quantile Networks for Distributional Reinforcement Learning
52917 A3C Noisy Networks for Exploration
47887 QR-DQN-1 Distributional Reinforcement Learning with Quantile Regression
47092 C51 A Distributional Perspective on Reinforcement Learning
42253 NoisyNet DuDQN Noisy Networks for Exploration
42214.0 DuDQN Dueling Network Architectures for Deep Reinforcement Learning
42120 QR-DQN-0 Distributional Reinforcement Learning with Quantile Regression
41324.5 PDD DQN Dueling Network Architectures for Deep Reinforcement Learning
37896.8 Distributional DQN Rainbow: Combining Improvements in Deep Reinforcement Learning
37275 DuDQN Noisy Networks for Exploration
35338.5 DDQN A Distributional Perspective on Reinforcement Learning
23633.0 DQN A Distributional Perspective on Reinforcement Learning
23633.0 DQN Rainbow: Combining Improvements in Deep Reinforcement Learning
20525 NoisyNet DQN Noisy Networks for Exploration
18688.9 Human Dueling Network Architectures for Deep Reinforcement Learning
18303 DQN Noisy Networks for Exploration
16667.5 IMPALA (deep, multitask) IMPALA: Scalable Distributed Deep-RL with Importance Weighted Actor-Learner Architectures
2874.5 Random Dueling Network Architectures for Deep Reinforcement Learning

Normal Starts

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