Track manipulation, locomotion and embodied systems.
Search papers, labs, and topics across Lattice.
We extend port-Hamiltonian neural networks (pHNNs) to model dynamic systems with multiple asymptotically stable equilibria by parametrizing the Hamiltonian as a product of Bregman divergences from an input-convex network. This overcomes the limitation of standard pHNNs which require a single global minimum in their Hamiltonian, thus restricting them to systems with only one attractor. The proposed method guarantees local Lyapunov stability and, under a hyperbolicity assumption, almost-everywhere stability, while empirically demonstrating improved convergence speeds on benchmark systems.
Standard port-Hamiltonian neural networks are fundamentally limited to single-attractor systems; this work unlocks multi-attractor dynamics by novelly parametrizing the Hamiltonian as a product of Bregman divergences.
Stable port-Hamiltonian neural networks certify asymptotic stability by construction. Yet, their Hamiltonian is a global Lyapunov function with a single global minimum, so they can represent only dynamic systems with {one} attractor. We demonstrate that this excludes even simple systems with energy landscapes forming a double well, and we overcome the restriction by parametrising the Hamiltonian as a {product} of Bregman divergences generated by one input-convex network. We prove that the resulting model is locally Lyapunov stable, that the coexistence of stable equilibria forces additional non-asymptotically-stable equilibria to exist, that all equilibria lie in a bounded region, and under a hyperbolicity assumption that almost-everywhere stability holds. On three systems our approach is able to recover the energy surface characteristics and improve the convergence speed by 1.8$\times$-8.5$\times$.