AlphaZero: Mixing in Dirichlet Noise
Beginner
The problem, from first principles
To keep self-play from repeating the same opening, AlphaZero mixes Dirichlet noise into the prior at the root of the search. The noise forces the search to consider moves the network would otherwise ignore.
From theory to code
Implement dirichlet_mix(P, noise, eps), blending the prior with the noise.
Constraints
epsis typically 0.25 in AlphaZero.
Hints
Hint 1
Weight the prior by 1 - eps and the noise by eps.
Keywords
research-papersreinforcement-learningself-playsearch
Python
Run your code to see results
No results yet
Run your code to check it against the examples, or submit to run every test.