predict method
Implementation
Map<String, Tensor> predict(Tensor state) {
// We use the specialized MuZero forward we added to the transformer
// If you haven't added the specific heads yet, this calls the logic:
// 1. Policy (using your lmHead/policyHead logic)
Tensor logits = transformer.lmHead.forward(state);
// 2. Value (Internal MLP logic)
Tensor vHidden = state.matmul(transformer.vW1);
for (int i = 0; i < vHidden.data.length; i++) {
if (vHidden.data[i] < 0) vHidden.data[i] = 0; // ReLU
}
Tensor value = vHidden.matmul(transformer.vW2);
for (int i = 0; i < value.data.length; i++) {
value.data[i] = _mathTanh(value.data[i]); // Tanh
}
return {"policy": logits, "value": value};
}