Overview

In Crazy Climber the player assumes the role of a person attempting to climb to the top of four skyscrapers. The climber is controlled via two joysticks.. There are a number of obstacles and dangers to avoid including:

  • Windows that open and close (the most common danger).
  • Bald-headed residents (a.k.a. Mad Doctors), who throw objects such as flower pots, buckets of water and fruit in an effort to knock the climber off the building (with larger objects appearing by more aggressive Mad Doctors in later levels).
  • A giant condor, who drops eggs and excrement aimed at the climber (two at a time in the early stages, four in later levels).
  • A giant ape (styled like King Kong), whose punch can prove deadly (he becomes more aggressive in later levels).
  • Falling steel girders and iron dumbbells (more numerous in the later levels).
  • Live wires, which protrude off electric signs.
  • Falling ‘Crazy Climber’ signs.

Some of these dangers appear at every level of the game; others make appearances only in later stages. Should the climber succumb to any one of these dangers, a new climber takes his place at the exact point where he fell; the last major danger is eliminated.

One ally the climber has is a pink “Lucky Balloon”; if he is able to grab it, the climber is transported up 8 stories to a window. The window onto which it drops the climber may be about to close. If the window that the climber is dropped onto is fully closed, the balloon pauses there until the window opens up again. The player does not actually earn bonus points for catching the balloon, but he is awarded the normal ‘step value’ for each of the eight floors that he passes while holding the balloon.

If the climber is able to ascend to the top of a skyscraper and grabs the runner of a waiting helicopter, he earns a bonus and is transported to another skyscraper, which presents more dangers than the past. The helicopter would only wait about 30 seconds, then fly off.

If the player completes all four skyscrapers, he is taken back to the first skyscraper and the game restarts from the beginning, but the player keeps his score.

The difficulty level of any game was modified to take into account the skill of previous players. Hence if a player pushed the high score up to say 250,000 (needed a really good player), any novice player following would get thoroughly wiped out for several games due to the increased difficulty level, and would have to play until it dropped back down.

Description from Wikipedia

State of the Art

Human Starts

Result Method Type Score from
263953.5 ApeX DQN DQN Distributed Prioritized Experience Replay
154416.5 DistributionalDQN DQN Rainbow: Combining Improvements in Deep Reinforcement Learning
143962.0 RainbowDQN DQN Rainbow: Combining Improvements in Deep Reinforcement Learning
138518.0 A3C LSTM PG Asynchronous Methods for Deep Learning
131086.0 PERDDQN (prop) DQN Prioritized Experience Replay
127853.0 DuelingPERDQN DQN Dueling Network Architectures for Deep Reinforcement Learning
127512.0 PERDDQN (rank) DQN Prioritized Experience Replay
124566.0 DuelingDDQN DQN Dueling Network Architectures for Deep Reinforcement Learning
113782.0 DDQN DQN Deep Reinforcement Learning with Double Q-learning
112646.0 A3C FF (4 days) PG Asynchronous Methods for Deep Learning
109337.0 PERDQN (rank) DQN Prioritized Experience Replay
101624.0 A3C FF (1 day) PG Asynchronous Methods for Deep Learning
98576.5 NoisyNetDQN DQN Rainbow: Combining Improvements in Deep Reinforcement Learning
98128.0 DQN2015 DQN Dueling Network Architectures for Deep Reinforcement Learning
65451.0 GorilaDQN DQN Massively Parallel Methods for Deep Reinforcement Learning
50992.0 DQN2015 DQN Massively Parallel Methods for Deep Reinforcement Learning
32667.0 Human Human Massively Parallel Methods for Deep Reinforcement Learning
9337.0 Random Random Massively Parallel Methods for Deep Reinforcement Learning

No-op Starts

Result Method Type Score from
320426.0 ApeX DQN DQN Distributed Prioritized Experience Replay
183137.0 PERDDQN (prop) DQN Rainbow: Combining Improvements in Deep Reinforcement Learning
179877.0 C51 Misc A Distributional Perspective on Reinforcement Learning
178355.0 DistributionalDQN DQN Rainbow: Combining Improvements in Deep Reinforcement Learning
171171.0 NoisyNet-DuelingDQN DQN Noisy Networks for Exploration
168788.5 RainbowDQN DQN Rainbow: Combining Improvements in Deep Reinforcement Learning
163335.0 DuelingDQN DQN Noisy Networks for Exploration
162224.0 DuelingPERDQN DQN Dueling Network Architectures for Deep Reinforcement Learning
151909.5 DQfD Imitation Deep Q-Learning from Demonstrations
150444.0 ACKTR PG Scalable trust-region method for deep reinforcement learning using Kronecker-factored approximation
143570.0 DuelingDDQN DQN Dueling Network Architectures for Deep Reinforcement Learning
141161.0 PER DQN Scalable trust-region method for deep reinforcement learning using Kronecker-factored approximation
141161.0 PERDDQN (rank) DQN Dueling Network Architectures for Deep Reinforcement Learning
139950.0 NoisyNet-A3C PG Noisy Networks for Exploration
136828.2 DuelingPERDDQN DQN Deep Q-Learning from Demonstrations
134783.0 A3C PG Noisy Networks for Exploration
119679.0 DDQN+PopArt DQN Learning values across many orders of magnitude
118768.0 NoisyNetDQN DQN Rainbow: Combining Improvements in Deep Reinforcement Learning
118305.0 NoisyNet-DQN DQN Noisy Networks for Exploration
117282.0 DDQN DQN Dueling Network Architectures for Deep Reinforcement Learning
116480.0 DQN DQN Noisy Networks for Exploration
114103 DQN2015 DQN Human-level control through deep reinforcement learning
110763.0 DQN2015 DQN Dueling Network Architectures for Deep Reinforcement Learning
101874.0 DDQN DQN Deep Reinforcement Learning with Double Q-learning
85919.16 GorilaDQN DQN Massively Parallel Methods for Deep Reinforcement Learning
35829.4 Human Human Dueling Network Architectures for Deep Reinforcement Learning
35411 Human Human Human-level control through deep reinforcement learning
23411 Linear Misc Human-level control through deep reinforcement learning
10781 Random Random Human-level control through deep reinforcement learning
10780.5 Random Random Learning values across many orders of magnitude
149.8 Contingency Misc Human-level control through deep reinforcement learning

Normal Starts

Result Method Type Score from
132461.0 ACER PG Proximal Policy Optimization Algorithms
110202.0 PPO PG Proximal Policy Optimization Algorithms
107770.0 A2C PG Proximal Policy Optimization Algorithms