AlphaZero: The PUCT Score
Advanced
The problem, from first principles
AlphaZero (Silver et al., 2017) guides Monte Carlo tree search with a neural network. Each candidate move gets a score that balances its measured value against the network's prior, with a bonus for moves that have not been visited much.
From theory to code
Implement puct_score(Q, P, N_parent, N_child, c), returning the value plus the exploration bonus.
Constraints
- The bonus uses the square root of the parent's visits.
Hints
Hint 1
Compute c * P * sqrt(N_parent) / (1 + N_child) and add it to Q.
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.