ACM AI at UCLA · Student project · Fall 2026–present
Scoring the available poker actions.Poker Decision Modeling
I implemented a MiniLM action scorer and contributed model-engine integration to the PocketSolver team prototype.
- Group
- ACM AI at UCLA
- Status
- In progress
- Output
- Fold, call, raise probabilities
The task
PocketSolver scores the available actions in a poker situation. My implemented model predicts reference actions from PokerBench contexts, questions, and candidate menus. Opponent-history adaptation is a possible extension, not an implemented feature.
My model
I implemented a PyTorch scorer using a 33M-parameter MiniLM encoder. It encodes context/question/candidate pairs in a batch, pools candidate tokens, applies a shared scalar head, and normalizes scores over the supplied actions.
Variable-length candidate lists use attention and candidate-token masks, with padded action slots excluded. The pipeline separates model inputs from targets and saves reloadable checkpoints, tokenizer, and configuration.
Data and evaluation
I audited 563,200 PokerBench source records for duplicates, conflicting labels, and related-scenario overlap. The modeling pool was 10,000 development examples, not the entire audited source.
I used scenario-grouped five-fold cross-validation and compared with uniform, menu-prior, and linear baselines. The official PokerBench test set was not used for these development results.
What the comparison showed
My MiniLM model achieved 65.20% out-of-fold exact-action accuracy and log loss 0.713931. The linear baseline achieved 69.69% accuracy and log loss 1.002538. The model’s log loss was 28.8% lower, while the baseline had higher action accuracy.
These are development comparisons, not evidence of profitable, unexploitable, or GTO play.
Team integration
I contributed to PocketSolver’s model-engine wrapper for side-by-side local-model and hosted-LLM predictions, with a shared probability-and-latency interface.
