Skip to main content
Back to timeline
arXivSource publication:

Structured neural Lyapunov functions guarantee positive definiteness by design, and PLNet reaches the largest certified region on the inverted pendulum

Synopsis

For known nonlinear discrete-time systems, a controller and a control Lyapunov function can be trained jointly while positive definiteness of the candidate no longer rests on penalties at sampled states. Input convex neural networks and Polyak-Łojasiewicz networks, which carry lower and upper quadratic bounds and are therefore positive definite by construction, satisfy positivity and radial unboundedness directly, leaving only the closed-loop decrease condition for formal verification. On an inverted pendulum and a path-tracking benchmark, the PLNet attains the largest base and expanded certified regions on the inverted pendulum, and all five seeds certify in a single round.

Source-provided article image: Learning and Verifying Stabilizing Controllers with Structured Neural Control Lyapunov Functions
arXiv

Interpretation

The controller and Lyapunov function are still trained jointly, but positive definiteness and radial unboundedness of the candidate now come directly from the network structure rather than from norm or quadratic penalties during training. Earlier approaches mostly encouraged positive definiteness through explicit norm or quadratic terms during training, then formally verified all Lyapunov conditions. On the inverted pendulum the PLNet attains the largest base and expanded certified regions of all architectures, and across five seeds both the PLNet and the quadratic baseline need only one certification round; ICNN-L2 is the only architecture that fails to certify the training region for one seed.

The verification step extends mixed-integer linear programming to mixed-integer quadratic programming so that networks with quadratic outputs can be handled, and it enlarges the certified region ring by ring without retraining. Earlier discrete-time neural Lyapunov control used mixed-integer linear programming verification, which does not directly accommodate quadratic Lyapunov functions. ICNN-L2, the PLNet and the quadratic baseline are verified with mixed-integer quadratic programming, while ICNN-L1, the MLP and MLP-ReLU-L1 use mixed-integer linear programming, all with CPLEX 22.2.0.0 and a 15-minute limit per run.

Training proceeds in three phases: the controller first imitates an LQR controller on the linearized system, the Lyapunov function is then trained alone, and finally both are trained together and passed to the verifier. Joint training of controller and Lyapunov function from random initialization often fails to converge, because a randomly initialized MLP controller is highly nonlinear and can make the closed loop unstable. Each benchmark evaluates six parameterizations over five seeds (11 to 15), with a shared 177-parameter MLP controller and PLNet, ICNN and MLP candidates of 342, 323 and 321 trainable parameters respectively.

Perspective

The demonstrated scope is two low-dimensional discrete-time benchmarks with known dynamics, an inverted pendulum and a path-tracking system, both with a 0.05-second sampling time, verified cell by cell over hyperrectangles partitioning the training domain. Verification deliberately excludes a small neighborhood of the origin, so it yields practical stability guarantees rather than asymptotic stability of the origin.

The ratio of the PLNet's monotonicity constant to its Lipschitz constant shapes the isotropy of the sublevel sets and the size of the certified region; the smallest tested ratio gives the largest region on both systems, but overly restrictive bounds may limit expressivity and prevent successful training or verification. On path tracking the quadratic baseline's maximum expanded region exceeds all neural candidates, so the structural advantage is not uniform across metrics.