Evaluating Model-Free Policy Optimization in Masked-Action Environments via an Exact Blackjack Oracle

Kevin Song · 2026

Read Evaluating Model-Free Policy Optimization in Masked-Action Environments via an Exact Blackjack Oracle by Kevin Song (2026) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2603-18642.txt.

Description

Song tests model-free policy learning on blackjack with an exact oracle as ground truth. Fixed Vegas rules define 4,600 decision cells with masked actions. Dynamic programming gives true action values and an expected value of minus 0.00161 per hand. Masked REINFORCE, SPSA, and the cross-entropy method train by simulation. REINFORCE learns fastest, matching 46.37 percent of optimal actions after one million hands. All three still show large regret in rare cells. Song also proves that bet sizing without counting collapses to the table minimum.

Catalog record

CREATORKevin Song
YEAR2026
COLLECTIONPAPERS
FORMATTXT
FILE SIZE23436
LICENSEhttp://creativecommons.org/licenses/by/4.0/
RIGHTSCC BY
SOURCEarXiv

Source: https://arxiv.org/abs/2603.18642

Browse the full library · About the Tower of Babel