Neuroevolution

Neuroevolution produces offspring through selection, crossover and mutation; in each generation the best-performing networks are retained and reproduced, gradually optimizing the strategy.

Neuroevolution does not rely on a differentiable loss function and is suitable for sparse-reward scenarios in reinforcement learning.

Neuroevolution treats network weights as genes, directly uses the final game score as fitness, and searches the parameter space through population-level selection, crossover and mutation.

This gradient-free approach of neuroevolution is more robust to environments with sparse, non-differentiable or noisy rewards, but usually requires greater computation to evaluate each individual in the population.

Neuroevolution can serve as an effective alternative or supplement to gradient methods in sparse-reward scenarios.

Human players are limited by reaction speed and attention, whereas neuroevolution agents can simulate thousands of attempts at high speed and discover shortcuts or tricks that humans find hard to notice.

Neuroevolution does not train a single agent, but lets a group of neural networks compete, mutate and evolve in the game, ultimately selecting the individuals with excellent performance.

Neuroevolution is a method of training neural networks with evolutionary algorithms.

Neuroevolution maintains a population of neural networks, runs each individual once in the game, performs selection, crossover and mutation based on the score, and eliminates poorly performing strategies generation by generation.

The project trains AI in a 3D Super Mario environment, with fitness determined by level progress and score.

The demo video provided by the project shows that the AI can perform operations such as continuous jumping, avoiding enemies and precise landing, with smoothness approaching or even exceeding that of ordinary human players.

Neuroevolution does not require a differentiable environment model and is more robust to sparse rewards and long-sequence decision-making.

The advantage of neuroevolution lies in its simple implementation and ease of parallelization—each individual in the population can run independently, making it naturally suited to multi-core or distributed training.

Some evolved strategies exploit subtle flaws in the game physics engine to achieve ultra-long jumps or fast level clears.

The combination of neuroevolution with other reinforcement learning methods may drive the development of more general agents.