NEAT
A neuroevolution method that simultaneously optimizes neural network topology and weights using evolutionary algorithms. Proposed by Kenneth Stanley and Risto Miikkulainen in 2002, it introduces speciation mechanisms and historical markers to enable automatic evolution of network structure.
Timeline (last 90 days)
Kenneth Stanley's NEAT algorithm and novelty search theory indicate that using objective fitness as the sole guiding metric actually impedes exploration.
NEAT was proposed by Kenneth Stanley and Risto Miikkulainen in 2002.
The core innovation of NEAT lies in the introduction of a speciation mechanism to protect the early survival of innovative structures.
NEAT uses historical markings to track gene homology, enabling meaningful crossover between networks of different topologies.
A Reddit developer used a Mario controller evolved with NEAT to successfully complete the level 《Bowser's Crown 4-1》.
Unlike traditional reinforcement learning, NEAT does not rely on gradient descent but instead simulates natural selection to iterate.
The backtracking and nonlinear complex layouts of ROM Hack levels are extremely unfriendly to networks evolved based on a reward signal of moving rightward.
Solving the backtracking problem typically requires more refined reward shaping or the introduction of temporal memory mechanisms, as standard NEAT's feedforward or simple recurrent networks have limitations in handling long-term temporal dependencies.
Each new level in the project is trained separately, and completing a level does not demonstrate that the network has generalization capability for unfamiliar levels.
NEAT is naturally well-suited for game environments with discrete action spaces and sparse rewards, and can automatically evolve network topology.
1 more timeline events
All Facts (10)
NEAT was proposed by Kenneth Stanley and Risto Miikkulainen in 2002.
60%UnverifiedKenneth Stanley's NEAT algorithm and novelty search theory indicate that using objective fitness as the sole guiding metric actually impedes exploration.
50%UnverifiedNEAT uses historical markings to track gene homology, enabling meaningful crossover between networks of different topologies.
50%UnverifiedIn the project, the neural network reads nearby game state information (enemies, terrain, Mario's position) from the emulator's memory and outputs button operations.
50%UnverifiedUnlike traditional reinforcement learning, NEAT does not rely on gradient descent but instead simulates natural selection to iterate.
50%UnverifiedSolving the backtracking problem typically requires more refined reward shaping or the introduction of temporal memory mechanisms, as standard NEAT's feedforward or simple recurrent networks have limitations in handling long-term temporal dependencies.
50%UnverifiedEach new level in the project is trained separately, and completing a level does not demonstrate that the network has generalization capability for unfamiliar levels.
50%UnverifiedNEAT is naturally well-suited for game environments with discrete action spaces and sparse rewards, and can automatically evolve network topology.
50%UnverifiedThe backtracking and nonlinear complex layouts of ROM Hack levels are extremely unfriendly to networks evolved based on a reward signal of moving rightward.
50%UnverifiedThe core innovation of NEAT lies in the introduction of a speciation mechanism to protect the early survival of innovative structures.
50%