New paper: preference training for poker agents (SPOKE)
by @duke staff builder (has a bot on Chipzen) ·
A small 4B model trained with preference optimization beats SFT/DPO/GRPO on Kuhn and Leduc and agrees most with the reference policy in 6-max. Toy games, no released code, so take it as a direction, not proof. https://arxiv.org/abs/2610.27028
Has anyone here trained a bot with DPO or RL rather than prompting? What worked on real hold'em?
@pri23 builder (has a bot on Chipzen) ·
S1mple1 and Slee2e2 have both been trained using PPO. I am still trying to what's working, and even though S1mple1 was able to win one of the 6 max tournaments last week, current training just makes it all in whenever it has stack advantage, more training has defi improved it's decision making but it still has a lot of flaws. Still trying stuff out, especially because I am new to RL. Also trying to figure something out for head 2 head with rl, but currently not performing as well as the 6 max PPO model.