Overview

The player controls an insect-like creature called a Yar who must nibble or shoot through a barrier in order to fire his Zorlon Cannon into the breach. The objective is to destroy the evil Qotile, which exists on the other side of the barrier. The Qotile can attack the Yar, even if the barrier is undamaged, by turning into the Swirl and shooting across the screen. In early levels the player is warned before the Swirl is fired, and he can retreat to a safe distance to dodge the attack. The Yar can hide from a pursuing destroyer missile within a “neutral zone” in the middle of the screen, but the Yar cannot shoot while in the zone. The Swirl can kill the Yar anywhere, even inside the Neutral Zone.

To destroy the Qotile or the Swirl, the player has to either touch the Qotile or eat a piece of the shield to activate the Zorlon Cannon, aim the cannon by leading the with the Qotile or Swirl, then fire the cannon and fly the Yar out of the path of the cannon’s shot. If the weapon finds its mark, the level ends. If the cannon blast hits a piece of the shield or misses, it is expended. The cannon itself is dangerous to the player, for once it is activated, the fire button launches it instead of firing the Yar’s usual shots, and as the cannon tracks the Yar’s vertical position, players effectively use the Yar itself as a target and therefore must immediately maneuver to avoid being hit by their own shot. The cannon shot can also rebound off the shield in later levels.

Description from Wikipedia

State of the Art

Human Starts

Result Method Type Score from
131701.1 ApeX DQN DQN Distributed Prioritized Experience Replay
93007.9 RainbowDQN DQN Rainbow: Combining Improvements in Deep Reinforcement Learning
58145.9 DuelingPERDQN DQN Dueling Network Architectures for Deep Reinforcement Learning
47135.2 Human Human Deep Reinforcement Learning with Double Q-learning
25976.5 DuelingDDQN DQN Dueling Network Architectures for Deep Reinforcement Learning
8267.7 DistributionalDQN DQN Rainbow: Combining Improvements in Deep Reinforcement Learning
7270.8 A3C FF (1 day) PG Asynchronous Methods for Deep Learning
7157.5 A3C FF (4 days) PG Asynchronous Methods for Deep Learning
6626.7 PERDQN (rank) DQN Prioritized Experience Replay
6270.6 DDQN DQN Deep Reinforcement Learning with Double Q-learning
5965.1 PERDDQN (prop) DQN Prioritized Experience Replay
5615.5 A3C LSTM PG Asynchronous Methods for Deep Learning
5487.3 NoisyNetDQN DQN Rainbow: Combining Improvements in Deep Reinforcement Learning
4687.4 PERDDQN (rank) DQN Prioritized Experience Replay
4577.5 DQN2015 DQN Dueling Network Architectures for Deep Reinforcement Learning
1476.9 Random Random Deep Reinforcement Learning with Double Q-learning
-9001.0 DQN2015 DQN Asynchronous Methods for Deep Learning
-9001.0 GorilaDQN DQN Asynchronous Methods for Deep Learning

No-op Starts

Result Method Type Score from
148594.8 ApeX DQN DQN Distributed Prioritized Experience Replay
125169.0 ACKTR PG Scalable trust-region method for deep reinforcement learning using Kronecker-factored approximation
102557.0 RainbowDQN DQN Rainbow: Combining Improvements in Deep Reinforcement Learning
86101.0 NoisyNet-DuelingDQN DQN Noisy Networks for Exploration
69618.1 DuelingPERDQN DQN Dueling Network Architectures for Deep Reinforcement Learning
63484.4 DuelingPERDDQN DQN Deep Q-Learning from Demonstrations
61755.0 NoisyNet-A3C PG Noisy Networks for Exploration
61575.7 DQfD Imitation Deep Q-Learning from Demonstrations
54576.9 Human Human Dueling Network Architectures for Deep Reinforcement Learning
49622.1 DuelingDDQN DQN Dueling Network Architectures for Deep Reinforcement Learning
43120.0 DuelingDQN DQN Noisy Networks for Exploration
35050.0 C51 Misc A Distributional Perspective on Reinforcement Learning
23915.0 NoisyNet-DQN DQN Noisy Networks for Exploration
21596.0 A3C PG Noisy Networks for Exploration
21409.5 DDQN+PopArt DQN Learning values across many orders of magnitude
20648.0 DQN DQN Noisy Networks for Exploration
18098.9 DQN2015 DQN Dueling Network Architectures for Deep Reinforcement Learning
18089.9 DQN2015 DQN Rainbow: Combining Improvements in Deep Reinforcement Learning
16608.6 DistributionalDQN DQN Rainbow: Combining Improvements in Deep Reinforcement Learning
16451.7 PERDDQN (prop) DQN Rainbow: Combining Improvements in Deep Reinforcement Learning
11712.6 DDQN DQN Dueling Network Architectures for Deep Reinforcement Learning
11357.0 PER DQN Scalable trust-region method for deep reinforcement learning using Kronecker-factored approximation
11357.0 PERDDQN (rank) DQN Dueling Network Architectures for Deep Reinforcement Learning
9570.1 NoisyNetDQN DQN Rainbow: Combining Improvements in Deep Reinforcement Learning
3092.9 Random Random Dueling Network Architectures for Deep Reinforcement Learning