Overview

Marooned on the ice planet Krybor, the player uses a laser cannon to destroy legions of demons that attack from above. Visually, the demons appear in waves similar to other space-themed shooters, but individually combine from the sides of the screen to the area above the player’s cannon.

Each wave introduces new weapons with which the demons attack, such as long streaming lasers and laser clusters. Starting in Wave 5, demons also divide into two smaller, bird-like creatures that eventually attempt descent onto the player’s cannon. Starting in Wave 9, the demons’ shots follow directly beneath the monsters, making it difficult for the player to slip underneath to get in a direct shot.

Description from Wikpedia

State of the Art

Human Starts

Result Method Type Score from
133002.1 ApeX DQN DQN Distributed Prioritized Experience Replay
115201.9 A3C LSTM PG Asynchronous Methods for Deep Learning
113308.4 A3C FF (4 days) PG Asynchronous Methods for Deep Learning
109856.6 DistributionalDQN DQN Rainbow: Combining Improvements in Deep Reinforcement Learning
109670.7 RainbowDQN DQN Rainbow: Combining Improvements in Deep Reinforcement Learning
84997.5 A3C FF (1 day) PG Asynchronous Methods for Deep Learning
73371.3 DuelingPERDQN DQN Dueling Network Architectures for Deep Reinforcement Learning
73185.8 PERDDQN (prop) DQN Prioritized Experience Replay
69803.4 DDQN DQN Deep Reinforcement Learning with Double Q-learning
61277.5 PERDDQN (rank) DQN Prioritized Experience Replay
56322.8 DuelingDDQN DQN Dueling Network Architectures for Deep Reinforcement Learning
25207.8 NoisyNetDQN DQN Rainbow: Combining Improvements in Deep Reinforcement Learning
19478.8 PERDQN (rank) DQN Prioritized Experience Replay
14880.13 GorilaDQN DQN Massively Parallel Methods for Deep Reinforcement Learning
12835.2 DQN2015 DQN Massively Parallel Methods for Deep Reinforcement Learning
12550.7 DQN2015 DQN Dueling Network Architectures for Deep Reinforcement Learning
3442.8 Human Human Massively Parallel Methods for Deep Reinforcement Learning
208.3 Random Random Massively Parallel Methods for Deep Reinforcement Learning

No-op Starts

Result Method Type Score from
274176.7 ACKTR PG Scalable trust-region method for deep reinforcement learning using Kronecker-factored approximation
133086.4 ApeX DQN DQN Distributed Prioritized Experience Replay
130955.0 C51 Misc A Distributional Perspective on Reinforcement Learning
111185.2 RainbowDQN DQN Rainbow: Combining Improvements in Deep Reinforcement Learning
110626.5 DistributionalDQN DQN Rainbow: Combining Improvements in Deep Reinforcement Learning
72878.6 DuelingPERDQN DQN Dueling Network Architectures for Deep Reinforcement Learning
71846.4 PER DQN Scalable trust-region method for deep reinforcement learning using Kronecker-factored approximation
71846.4 PERDDQN (rank) DQN Dueling Network Architectures for Deep Reinforcement Learning
70171.8 PERDDQN (prop) DQN Rainbow: Combining Improvements in Deep Reinforcement Learning
69311.0 NoisyNet-DuelingDQN DQN Noisy Networks for Exploration
63644.9 DDQN+PopArt DQN Learning values across many orders of magnitude
61033.0 DuelingDQN DQN Noisy Networks for Exploration
60813.3 DuelingDDQN DQN Dueling Network Architectures for Deep Reinforcement Learning
58044.2 DDQN DQN Dueling Network Architectures for Deep Reinforcement Learning
37880.0 NoisyNet-A3C PG Noisy Networks for Exploration
37085.0 A3C PG Noisy Networks for Exploration
36150.0 NoisyNet-DQN DQN Noisy Networks for Exploration
24950.1 NoisyNetDQN DQN Rainbow: Combining Improvements in Deep Reinforcement Learning
13693.12 GorilaDQN DQN Massively Parallel Methods for Deep Reinforcement Learning
12696.0 DQN DQN Noisy Networks for Exploration
12149.4 DQN2015 DQN Dueling Network Architectures for Deep Reinforcement Learning
9711.9 DDQN DQN Deep Reinforcement Learning with Double Q-learning
9711 DQN2015 DQN Human-level control through deep reinforcement learning
3848.8 DQfD Imitation Deep Q-Learning from Demonstrations
3511.6 DuelingPERDDQN DQN Deep Q-Learning from Demonstrations
3401 Human Human Human-level control through deep reinforcement learning
1971.0 Human Human Dueling Network Architectures for Deep Reinforcement Learning
520.5 Linear Misc Human-level control through deep reinforcement learning
152.1 Random Random Human-level control through deep reinforcement learning
0 Contingency Misc Human-level control through deep reinforcement learning

Normal Starts

Result Method Type Score from
38808.3 ACER PG Proximal Policy Optimization Algorithms
11378.4 PPO PG Proximal Policy Optimization Algorithms
6639.1 A2C PG Proximal Policy Optimization Algorithms