Overview

The bottom two thirds of the screen are covered by a mass of water with four rows of ice blocks floating horizontally. The player moves by jumping from one row to another while trying to avoid various kinds of foes including crabs and birds. There are also fish which grant extra points.

On the top of the screen is the shore where the player must build the Igloo. From the fourth level onwards there is also a polar bear walking around on the shore which must be avoided.

The game levels alternate between large ice blocks and little ice pieces. The levels with the little pieces are actually easier, since one can walk left or right over them without falling in the water.

Each time the player jumps on a piece of ice in a row its color changes from white to blue and the player gets an ice block in the Igloo on the shore. The player has the ability to change the direction in which the ice is flowing by pressing the fire button, but that costs a piece of the Igloo.

After the player has jumped on all the pieces on the screen, they all turn back to white and one can jump on them again. When all the 15 ice blocks required for building the Igloo are gathered, the player has to get back to the shore and get inside it, thus proceeding to the next level. On every level the enemies and the ice blocks move slightly faster than in the previous level making the game more difficult.

Each level must be completed in 45 seconds, (represented as the declining temperature,) else the eskimo dies frozen. The faster the level is completed the more bonus points are awarded to the player. If player makes it past level 20, a “magic” fish will appear between the temperature gage and the number of lives remaining, this serves no real purpose other than as an Easter egg to the game.

Description from Wikipedia

State of the Art

Human Starts

Result Method Type Score from
6511.5 ApeX DQN DQN Distributed Prioritized Experience Replay
4202.8 Human Human Massively Parallel Methods for Deep Reinforcement Learning
4141.1 RainbowDQN DQN Rainbow: Combining Improvements in Deep Reinforcement Learning
4038.4 DuelingPERDQN DQN Dueling Network Architectures for Deep Reinforcement Learning
3510.0 PERDDQN (rank) DQN Prioritized Experience Replay
2930.2 PERDDQN (prop) DQN Prioritized Experience Replay
2813.9 DistributionalDQN DQN Rainbow: Combining Improvements in Deep Reinforcement Learning
2332.4 DuelingDDQN DQN Dueling Network Architectures for Deep Reinforcement Learning
1448.1 DDQN DQN Deep Reinforcement Learning with Double Q-learning
496.1 DQN2015 DQN Dueling Network Architectures for Deep Reinforcement Learning
426.6 GorilaDQN DQN Massively Parallel Methods for Deep Reinforcement Learning
418.8 NoisyNetDQN DQN Rainbow: Combining Improvements in Deep Reinforcement Learning
288.7 PERDQN (rank) DQN Prioritized Experience Replay
197.6 A3C LSTM PG Asynchronous Methods for Deep Learning
190.5 A3C FF (4 days) PG Asynchronous Methods for Deep Learning
180.1 A3C FF (1 day) PG Asynchronous Methods for Deep Learning
157.4 DQN2015 DQN Massively Parallel Methods for Deep Reinforcement Learning
66.4 Random Random Massively Parallel Methods for Deep Reinforcement Learning

No-op Starts

Result Method Type Score from
9590.5 RainbowDQN DQN Rainbow: Combining Improvements in Deep Reinforcement Learning
9328.6 ApeX DQN DQN Distributed Prioritized Experience Replay
7413.0 DuelingPERDQN DQN Dueling Network Architectures for Deep Reinforcement Learning
4672.8 DuelingDDQN DQN Dueling Network Architectures for Deep Reinforcement Learning
4380.1 PERDDQN (rank) DQN Dueling Network Architectures for Deep Reinforcement Learning
4335 Human Human Human-level control through deep reinforcement learning
3965.0 C51 Misc A Distributional Perspective on Reinforcement Learning
3938.2 DistributionalDQN DQN Rainbow: Combining Improvements in Deep Reinforcement Learning
3469.6 DDQN+PopArt DQN Learning values across many orders of magnitude
3421.6 PERDDQN (prop) DQN Rainbow: Combining Improvements in Deep Reinforcement Learning
2923.0 NoisyNet-DuelingDQN DQN Noisy Networks for Exploration
2807.0 DuelingDQN DQN Noisy Networks for Exploration
1683.3 DDQN DQN Dueling Network Architectures for Deep Reinforcement Learning
1000.0 DQN DQN Noisy Networks for Exploration
797.4 DQN2015 DQN Dueling Network Architectures for Deep Reinforcement Learning
753.0 NoisyNet-DQN DQN Noisy Networks for Exploration
605.16 GorilaDQN DQN Massively Parallel Methods for Deep Reinforcement Learning
583.6 NoisyNetDQN DQN Rainbow: Combining Improvements in Deep Reinforcement Learning
328.3 DQN2015 DQN Human-level control through deep reinforcement learning
288.0 A3C PG Noisy Networks for Exploration
261.0 NoisyNet-A3C PG Noisy Networks for Exploration
241.5 DDQN DQN Deep Reinforcement Learning with Double Q-learning
216.9 Linear Misc Human-level control through deep reinforcement learning
180.9 Contingency Misc Human-level control through deep reinforcement learning
65.2 Random Random Human-level control through deep reinforcement learning

Normal Starts

Result Method Type Score from
314.2 PPO PG Proximal Policy Optimization Algorithms
285.6 ACER PG Proximal Policy Optimization Algorithms
261.8 A2C PG Proximal Policy Optimization Algorithms