AlphaZero Training Run: 62-Iteration Self-Play Metrics

• Policy loss dropped 67% (3.86 → 1.29) over 62 iterations, with most improvement in the first 15 iterations
• Value loss fell 96% (0.32 → 0.014) and fully converged by iteration 30
• Average game length shrank from 354 to 117 moves — a 67% reduction — as play became more decisive
• Policy-MCTS agreement nearly doubled (31% → 59%), though the network still overrides tree search 41% of the time
• Late-game value error improved from 0.72 to 0.39, but early-game error worsened (0.70 → 0.88) — the network struggles to evaluate opening positions

• Policy loss fell 67% (3.86 → 1.29) over 62 iterations with train and validation tracking closely — no overfitting
• Value loss dropped 96% (0.32 → 0.01), converging near zero by iteration 30
• Average game length collapsed from 354 to 117 moves as the agent stopped drifting into long, inconclusive games
• Policy agreement with MCTS nearly doubled (31% → 59%), but at 59% the network still frequently overrides the tree search
• Early-position value error rose from 0.70 to 0.88 even as late-game error fell — the agent struggles to evaluate openings

The value network's early-game error rose from 0.70 to 0.88 even as all other metrics improved. This suggests the network has learned strong tactical play but poor positional judgment in openings — a common pattern in self-play systems that lack opening books.