feat: enhance training and evaluation processes with new parameters and normalization options

CLAUDE.md

@@ -19,6 +19,7 @@ Try to stick to these commands for consistency.
 | `api`  | Model / shared types | (none) |
 | `core` | Primary business logic | api, rule |
 | `rule` | Game rules | api |
+| `bot`  | Bots and AI | api,rule,io |
 | `io`   | Export formats | api, core |
 | `ui`   | Entrypoint & UI | core, io |
 

modules/bot/src/main/scala/de/nowchess/bot/bots/nnue/NNUE.scala

@@ -133,8 +133,15 @@ class NNUE:
     for j <- 0 until 32 do
       output += l2Output(j) * l3Weights(j)
 
-    // Convert from sigmoid(output) back to centipawns (output is trained as sigmoid(eval/400))
+    // Convert from tanh-normalized output back to centipawns
-    // Inverse sigmoid: eval/400 = ln(output / (1 - output))
+    // Training uses: eval_normalized = tanh(eval_cp / 300)
-    // But for simplicity, just scale directly: output ≈ sigmoid(eval/400), so eval ≈ 400 * (output - 0.5) * 2
+    // Inverse: eval_cp = 300 * atanh(output)
-    val cp = (output * 400f).toInt
+    // atanh(x) = 0.5 * ln((1 + x) / (1 - x))
+    val cp = if math.abs(output) >= 0.9999f then
+      // Clamp for numerical stability (avoid ln of very small numbers)
+      if output > 0f then 20000 else -20000
+    else
+      val atanh = 0.5f * math.log((1f + output) / (1f - output)).toFloat
+      (300f * atanh).toInt
+
     math.max(-20000, math.min(20000, cp))