scoring.py

"""
Reward and scoring utilities for Poker44 poker bot detection.
"""

from __future__ import annotations

from typing import Dict, List, Sequence

import numpy as np
import torch
from sklearn.metrics import average_precision_score, confusion_matrix, f1_score


def reward(y_pred: np.ndarray, y_true: np.ndarray) -> tuple[float, dict]:
    """
    Compute a reward based on F1, average precision and false-positive control.
    """
    preds = np.round(y_pred).astype(int)
    cm = confusion_matrix(y_true, preds, labels=[0, 1])
    _, fp, _, tp = cm.ravel()
    f1 = f1_score(y_true, preds) if (tp + fp) > 0 else 0.0
    ap_score = average_precision_score(y_true, y_pred) if y_pred.size else 0.0

    res = {
        "fp_score": 1 - fp / max(len(y_pred), 1),
        "f1_score": f1,
        "approve": ap_score
    }
    rew = sum(res.values()) / len(res)
    return rew, res