Atari Skiing Environment

Overview

Skiing is a single player only game, in which the player uses the joystick to control the direction and speed of a stationary skier at the top of the screen, while the background graphics scroll upwards, thus giving the illusion the skier is moving. The player must avoid obstacles, such as trees and moguls. The game cartridge contains five variations each of two principal games.

In the downhill mode, the player’s goal is to reach the bottom of the ski course as rapidly as possible, while a timer records his relative success.

In the slalom mode, the player must similarly reach the end of the course as rapidly as he can, but must at the same time pass through a series of gates (indicated by a pair of closely spaced flagpoles). Each gate missed counts as a penalty against the player’s time.

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
-3686.6 Human Deep Reinforcement Learning with Double Q-learning
-3686.6 Human Dueling Network Architectures for Deep Reinforcement Learning
-10169.1 Prioritized DDQN (rank, tuned) Prioritized Experience Replay
-10852.8 Prioritized DDQN (prop, tuned) Prioritized Experience Replay
-10911.1 A3C FF Asynchronous Methods for Deep Reinforcement Learning
-11490.4 DDQN (tuned) Deep Reinforcement Learning with Double Q-learning
-11685.8 Rainbow Rainbow: Combining Improvements in Deep Reinforcement Learning
-11928.0 DuDQN Dueling Network Architectures for Deep Reinforcement Learning
-12142.1 DQN Rainbow: Combining Improvements in Deep Reinforcement Learning
-13247.7 Distributional DQN Rainbow: Combining Improvements in Deep Reinforcement Learning
-13700.0 A3C FF 1 day Asynchronous Methods for Deep Reinforcement Learning
-14863.8 A3C LSTM Asynchronous Methods for Deep Reinforcement Learning
-15287.4 Random Deep Reinforcement Learning with Double Q-learning
-18955.8 PDD DQN Dueling Network Architectures for Deep Reinforcement Learning
-29404.3 DDQN Deep Reinforcement Learning with Double Q-learning
-29404.3 Prioritized DQN (rank) Prioritized Experience Replay

No-op Starts

Result Algorithm Source
-4336.9 Human Dueling Network Architectures for Deep Reinforcement Learning
-7550 NoisyNet DuDQN Noisy Networks for Exploration
-7989 DuDQN Noisy Networks for Exploration
-8857.4 DDQN A Distributional Perspective on Reinforcement Learning
-8857.4 DuDQN Dueling Network Architectures for Deep Reinforcement Learning
-8988.0 IMPALA (deep, multitask) IMPALA: Scalable Distributed Deep-RL with Importance Weighted Actor-Learner Architectures
-9163 QR-DQN-0 Distributional Reinforcement Learning with Quantile Regression
-9289 IQN Implicit Quantile Networks for Distributional Reinforcement Learning
-9324 QR-DQN-1 Distributional Reinforcement Learning with Quantile Regression
-10180.38 IMPALA (deep) IMPALA: Scalable Distributed Deep-RL with Importance Weighted Actor-Learner Architectures
-10632.9 Reactor ND The Reactor: A fast and sample-efficient Actor-Critic agent for Reinforcement Learning
-10753.4 Reactor The Reactor: A fast and sample-efficient Actor-Critic agent for Reinforcement Learning
-10870.6 Reactor The Reactor: A fast and sample-efficient Actor-Critic agent for Reinforcement Learning
-12630 DQN Noisy Networks for Exploration
-12957.8 Rainbow Rainbow: Combining Improvements in Deep Reinforcement Learning
-12972 A3C Noisy Networks for Exploration
-13062.3 DQN A Distributional Perspective on Reinforcement Learning
-13062.3 DQN Rainbow: Combining Improvements in Deep Reinforcement Learning
-13901 C51 A Distributional Perspective on Reinforcement Learning
-14763 NoisyNet DQN Noisy Networks for Exploration
-14959.8 Distributional DQN Rainbow: Combining Improvements in Deep Reinforcement Learning
-15970 NoisyNet A3C Noisy Networks for Exploration
-17098.1 Random Dueling Network Architectures for Deep Reinforcement Learning
-19949.9 PDD DQN Dueling Network Architectures for Deep Reinforcement Learning
-29975.0 IMPALA (shallow) IMPALA: Scalable Distributed Deep-RL with Importance Weighted Actor-Learner Architectures

Normal Starts

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