Value Transforms

inverse_transform(; epsilon=1e-10)

Create the inverse transform: loss → 1 / (loss + epsilon).

Keyword Arguments

• epsilon::Float64=1e-10: Small constant to avoid division by zero.

negation_transform()

Create the simple negation transform: loss → -loss.

This is the classic transform used in early MCTS implementations. Unbounded output range.

sigmoid_transform(; scale=1.0, center=0.0)

Create a sigmoid value transform: loss → 1 / (1 + exp(scale * (loss - center))).

Low loss maps to high value (~1), high loss maps to low value (~0). The center parameter shifts the midpoint (where output = 0.5). The scale parameter controls the steepness of the transition.

Keyword Arguments

• scale::Float64=1.0: Steepness of the sigmoid. Larger = sharper transition.
• center::Float64=0.0: Loss value at which the output is 0.5.

tanh_transform(; scale=1.0, center=0.0)

Create a tanh value transform: loss → (1 - tanh(scale * (loss - center))) / 2.

Rescaled to [0, 1] range. Low loss maps to high value (~1), high loss maps to low value (~0).

Keyword Arguments

• scale::Float64=1.0: Steepness of the transition.
• center::Float64=0.0: Loss value at which the output is 0.5.