Machine Learning and Reinforcement Learning
Learning objectiveUnderstand training data, rewards, evaluation, and the limits of small samples.
Understand the idea
Supervised learning can predict a label from observed features such as position, pot odds, board texture, and action history. It learns from recorded examples, including their biases. Reinforcement learning instead evaluates actions by a reward, often chip outcome, across repeated play. Exploration tries alternatives; exploitation uses the current policy. Self-play can generate experience but may learn artifacts of the simulated opponents.
Training belongs outside this website; the Academy adds no model service. Split data by time and opponent, prevent future cards from leaking into earlier decisions, and evaluate on unseen hands. A few dozen hands cannot establish a reliable improvement, especially with high poker variance. The production bot still must emit a legal, timely action.
At the poker table
A model trained with river cards in preflop features appears excellent in testing but is unusable live because those cards were unknown when the action occurred.
Common mistake
Do not use raw bankroll change as the sole reward without separating buy-ins, season resets, and variance.
Try it yourself
Design a feature list containing only information the bot knows preflop and explain how you would test on later hands.