Equivalent and Compact Representations of Neural Network Controllers With Decision Trees

Neural-network controllers can perform well while remaining difficult to inspect or verify. This work constructs soft decision-tree controllers that are exactly equivalent to discrete-output neural controllers, including networks with ReLU activations and argmax operations.

Why it matters

The transformation preserves the controller’s input-output behavior, so properties established on the tree apply to the original neural controller. Automatic branch pruning also makes the resulting representation more useful for solver-based analysis and implementation.

Key contributions

• Establishes an exact equivalence result for discrete-output neural-network and soft decision-tree controllers.
• Develops an automated transformation that removes redundant branches during construction.
• Evaluates the method on MountainCar, CartPole, and CarRacing controllers.
• Reports verification-time improvements of up to 21x on MountainCar and 2x on CartPole.