Summary Genetic Algorithms for QWOP Gait Evolution arxiv.org
6,716 words - PDF document - View PDF document
One Line
Genetic algorithms and keystrokes were used to optimize gaits for QWOP, with the help of dynamic mutation and the cellular model to enhance performance.
Slides
Slide Presentation (10 slides)
Key Points
- Genetic algorithms are used to discover effective gaits for the game QWOP.
- The authors developed a program called Totter to play the game programmatically.
- Different representations for the gaits were tested, with keystroke sequences being the most effective.
- Seeding the initial population with better-than-random individuals improved the performance of the genetic algorithms.
- Dynamic parameter control mechanisms were investigated, with dynamic mutation having a positive effect on performance.
- The cellular model of genetic algorithms consistently yielded stable and fast gaits.
- The evolved gaits did not reach the level of proficiency achieved by the best human players.
- Using a surrogate fitness function or faster evaluation times could improve the efficiency of the genetic algorithms.
Summaries
24 word summary
Genetic algorithms were used to find effective gaits for QWOP. Keystrokes were the most effective representation. Dynamic mutation and the cellular model improved performance.
77 word summary
This paper explores using genetic algorithms to discover effective gaits for the game QWOP. The authors developed a program called Totter to play the game programmatically. Different representations for the gaits were tested, with keystrokes being the most effective. Seeding with better-than-random individuals improved performance. Dynamic mutation had a positive effect. The cellular model consistently yielded stable and fast gaits. The authors suggest that dynamic sensory feedback would improve control. Limitations in fitness evaluation time were acknowledged.
163 word summary
This paper explores the use of genetic algorithms to discover effective gaits for the game QWOP. The authors developed a program called Totter to play the game programmatically, using the selenium library to control the mouse and keyboard. Different representations for the gaits were tested, and the keystroke representation was found to be the most effective. Seeding the initial population with better-than-random individuals greatly improved the performance of the genetic algorithms. Dynamic parameter control mechanisms were investigated, with dynamic mutation having a positive effect on performance. The cellular model consistently yielded stable and fast gaits. The authors were able to evolve gaits similar to those used by human players, but not at the same proficiency level. They suggest that dynamic sensory feedback would improve control. Limitations due to fitness evaluation time were acknowledged, and suggestions for improvement were made. Overall, this paper demonstrates the effectiveness of genetic algorithms for evolving gaits in QWOP and provides insights into various strategies and representations to use.
421 word summary
This paper explores the use of genetic algorithms to discover effective gaits for the game QWOP. QWOP is a difficult browser-based game where the player controls an Olympic sprinter using the 'Q', 'W', 'O', and 'P' keys.
The authors developed a program called Totter to play the game programmatically. Totter uses the selenium library to control the mouse and keyboard. The program translates the genotype of an individual into a phenotype, which is a sequence of key presses, holds, and releases. The fitness of each individual is evaluated by running their phenotype in the game and measuring the distance run.
Different representations for the gaits were tested, including keystroke sequences, keyup-keydown sequences, bitmask sequences, and bitmask-duration sequences. The keystroke representation, which encodes a sequence of key presses, was found to be the most effective in evolving stable and fast gaits. The other representations, allowing for more complex control patterns, did not perform as well.
The authors experimented with different initialization strategies and found that seeding the initial population with better-than-random individuals greatly improved the performance of the genetic algorithms. A seeding pool size of 500 was sufficient to achieve good results.
Dynamic parameter control mechanisms were investigated. Varying the mutation rate and selection pressure over time encouraged exploration in the early stages and convergence in the later stages. Dynamic mutation had a positive effect on performance, but dynamic replacement did not yield better results compared to static parameters.
Different types of genetic algorithms were compared, including steady-state, generational, and cellular models. The cellular model consistently yielded stable and fast gaits with low variability across trials. The steady-state model also performed well but had higher variability in fitness across trials.
The authors were able to evolve gaits similar to those used by human players of QWOP. However, the gaits evolved by the genetic algorithms did not reach the level of proficiency achieved by the best human players. The authors suggest that a system with dynamic sensory feedback would have a better chance of achieving human-like control of the runner.
The authors acknowledge limitations in their experiments due to the time required for fitness evaluation. They suggest that using a surrogate fitness function or building a simulation of QWOP with faster evaluation times could greatly improve the efficiency of the genetic algorithms.
In conclusion, this paper demonstrates the effectiveness of genetic algorithms for evolving gaits in the game QWOP. The authors provide insights into the best representations, initialization strategies, dynamic parameter control mechanisms, and types of genetic algorithms to use for this problem.
483 word summary
In this paper, the authors explore the use of genetic algorithms to automatically discover effective gaits for the game QWOP. QWOP is a browser-based game where the player controls an Olympic sprinter using the 'Q', 'W', 'O', and 'P' keys. The goal is to advance the runner to the end of the 100-meter race as quickly as possible. The game is known for its difficulty and unintuitive gameplay.
To play the game programmatically, the authors developed a program called Totter, which uses the selenium library to control the mouse and keyboard. The program translates the genotype of an individual into a phenotype, which is a sequence of key presses, holds, and releases. The fitness of each individual is evaluated by running their phenotype in the game and measuring the distance run.
The authors tested different representations for the gaits, including keystroke sequences, keyup-keydown sequences, bitmask sequences, and bitmask-duration sequences. They found that the keystroke representation, which simply encodes a sequence of key presses, was the most effective in evolving stable and fast gaits. The other representations, which allowed for more complex control patterns, did not perform as well.
The authors also experimented with different initialization strategies. They found that seeding the initial population with better-than-random individuals greatly improved the performance of the genetic algorithms. They tested different sizes of seeding pools and found that a pool size of 500 was sufficient to achieve good results.
In addition, the authors investigated dynamic parameter control mechanisms. They varied the mutation rate and selection pressure over time to encourage exploration in the early stages of evolution and convergence in the later stages. They found that dynamic mutation had some positive effect on performance, but dynamic replacement did not yield better results compared to static parameters.
Finally, the authors compared different types of genetic algorithms, including steady-state, generational, and cellular models. They found that the cellular model consistently yielded stable and fast gaits with low variability across trials. The steady-state model also performed well, but had higher variability in fitness across trials.
Overall, the authors were able to evolve gaits that were similar to those used by human players of QWOP. However, the gaits evolved by the genetic algorithms did not reach the level of proficiency achieved by the best human players. The authors suggest that a system with dynamic sensory feedback would have a better chance of achieving human-like control of the runner.
The authors acknowledge that their experiments were limited by the time required for fitness evaluation. They suggest that using a surrogate fitness function or building a simulation of QWOP with faster evaluation times could greatly improve the efficiency of the genetic algorithms.
In conclusion, this paper demonstrates the effectiveness of genetic algorithms for evolving gaits in the game QWOP. The authors provide insights into the best representations, initialization strategies, dynamic parameter control mechanisms, and types of genetic algorithms to use for this problem.