Evaluating Variance-Aware PUCT in Neural Chess Engines
(2026) EDAN70 20261Department of Computer Science
- Abstract (Swedish)
- Game-playing algorithms often face a trade-off between analyzing moves that currently look promising and exploring alternatives whose value is uncertain. AlphaZero-style chess engines combine neural network evaluation with Monte Carlo tree search to selectively explore the game tree. We evaluate a variance-aware modification of the PUCT tree policy on chess puzzles and find that it underperforms standard PUCT across different networks and search budgets under certain conditions. We suggest explanations for this discrepancy and possible improvements to the implementation of the variance-aware tree policy.
Please use this url to cite or link to this publication:
https://lup.lub.lu.se/student-papers/record/9249101
- author
- Ahlqvist, Axel LU and Näslund Cuesta, Oscar LU
- supervisor
- organization
- course
- EDAN70 20261
- year
- 2026
- type
- L3 - Miscellaneous, Projetcs etc.
- subject
- keywords
- Monte Carlo Tree Search, PUCT, chess engine, AlphaZero, reinforcement learning
- language
- English
- id
- 9249101
- date added to LUP
- 2026-08-26 10:40:24
- date last changed
- 2026-08-26 10:50:32
@misc{9249101,
abstract = {{Game-playing algorithms often face a trade-off between analyzing moves that currently look promising and exploring alternatives whose value is uncertain. AlphaZero-style chess engines combine neural network evaluation with Monte Carlo tree search to selectively explore the game tree. We evaluate a variance-aware modification of the PUCT tree policy on chess puzzles and find that it underperforms standard PUCT across different networks and search budgets under certain conditions. We suggest explanations for this discrepancy and possible improvements to the implementation of the variance-aware tree policy.}},
author = {{Ahlqvist, Axel and Näslund Cuesta, Oscar}},
language = {{eng}},
note = {{Student Paper}},
title = {{Evaluating Variance-Aware PUCT in Neural Chess Engines}},
year = {{2026}},
}