Deconstructing the Inductive Biases of Hamiltonian Neural Networks
This addresses the challenge of using physics-inspired models in practical, non-conservative systems such as robotics and reinforcement learning, representing an incremental improvement by relaxing biases for broader applicability.
The paper tackled the problem of applying physics-inspired neural networks to non-conservative systems like robotics by deconstructing their inductive biases, showing that improved generalization stems from modeling acceleration directly rather than symplectic structure, and achieved performance matching or exceeding on energy-conserving systems while dramatically improving on non-conservative ones.
Physics-inspired neural networks (NNs), such as Hamiltonian or Lagrangian NNs, dramatically outperform other learned dynamics models by leveraging strong inductive biases. These models, however, are challenging to apply to many real world systems, such as those that don't conserve energy or contain contacts, a common setting for robotics and reinforcement learning. In this paper, we examine the inductive biases that make physics-inspired models successful in practice. We show that, contrary to conventional wisdom, the improved generalization of HNNs is the result of modeling acceleration directly and avoiding artificial complexity from the coordinate system, rather than symplectic structure or energy conservation. We show that by relaxing the inductive biases of these models, we can match or exceed performance on energy-conserving systems while dramatically improving performance on practical, non-conservative systems. We extend this approach to constructing transition models for common Mujoco environments, showing that our model can appropriately balance inductive biases with the flexibility required for model-based control.