ReLU Networks for Model Predictive Control: Network Complexity and Performance Guarantees
Authors: Xingchen Li, Keyou You
Organizations: Department of Automation, and Beijing National Research Center for Info. Sci. & Tech. (BNRist), Tsinghua University, Beijing 100084, China.
Recent years have witnessed a resurgence in using ReLU neural networks (NNs) to represent model predictive control (MPC) policies. However, determining the sufficient network complexity to ensure closed-loop guarantees remains a fundamental open problem. This involves a critical precision-complexity trade-off: undersized networks may fail to capture the MPC policy, while oversized ones may outweigh the benefits of ReLU network approximation. In this work, we project the MPC actions onto a tightened constraint set to satisfy the hard constraints, and establish a state-dependent Lipschitz continuity property for the optimal MPC cost function, which enables sharp convergence analysis of the closed-loop system. For the first time, we derive explicit bounds on ReLU network width and depth for approximating MPC policies with closed-loop guarantees. To reduce network complexity for the convergence to user-specified target sets, we propose a non-uniform error framework with a state-aware scaling function to adaptively adjust both the input and output of the ReLU network. Our contributions provide a foundational step toward certifiable ReLU NN-based MPC.
Figures & tables
Fig. 1: Illustration of the main challenges and our analytical approach.
Fig. 2: Combinations of nw′ and nd′ satisfying ( 19 ).
Fig. 3: The state-aware scaling approach. The definitions of umpc(⋅) and unn(⋅) are given in ( 28 ) and ( 24 ), respectively.
Fig. 4: An example of β(x)=δ2(x)/δˉ and the input scaling function T(x) .
Fig. 5: An example of umpc(⋅) , umpc(⋅) and unn(⋅) . The shaded area in subfigure (c) represents the admissible error bound δ2(x) .
∥A∥
1.618
γ
185.9
c3
1.000
∥B∥
1.118
D(Xinv)
10.95
c4
56.84
d(U)
1
d(Xinv)
3.426
c5
246.9
D(U)
1
c0
70.39
c6
7.652×105
Lu
1.412
c1
1.069
c7
1.143×10−3
γmax
187.7
c2
15.83
δˉ
7.793×10−3
TABLE I: Constants of Theorem 1 for the system ( 34 ), with γ=0.99γmax .
Fig. 6: Sketch of the oscillating masses system [ 48 , 24 ] , with the damping not shown.
Fig. 7: Closed loop of the certified network for the system ( 34 ): 60 trajectories (blue) started at 0.97 of the boundary of Xinv (black), the set V(c6δ12) of Lemma 3 (red dashed), and the region on which the MPC input is saturated (grey).
Fig. 8: Decay of vmpc(xk) along the 60 trajectories of Fig. 7 (blue), the bound c6δ12+qkγ of Lemma 3 (dashed) and the level c6δ12 (dotted).
Fig. 9: Error ∥unn(x)−umpc(x)∥ of the two networks on the test states against ∥x∥ , for d=0.1 and d=0.001 . The dashed line is the constant bound δ of Theorem 1 , the solid line the bound δ2(x) of Theorem 2 .
Fig. 10: Closed loops of the two networks from 40 random states at 0.97 of the boundary of Xinv , with the input saturated to U : vmpc(xk) against k for d=0.1 and d=0.001 . The dotted line is the level c1d2 , whose sublevel set lies in the target set B(d) .
Fig. 11: The recursive sequence ak and its convergence behavior. The black curve is y=ψ(x) , the yellow line is the diagonal y=x , and the green line is the tangent of ψ(⋅) at its fixed point a∞ .
Model predictive control is a high-performance control method, but it incurs comparatively high online computational costs across many control tasks and is challenging to realize in low-level hardware settings. To address this, we present a framework for generating neural network controllers that provably inherit the safety and stability guarantees of a robust model predictive policy. In contrast to approaches based on post-verification or additional online safety filters, our method provides constructive, a priori design conditions that ensure preservation of recursive feasibility and stability before deployment. We treat the approximation error as a bounded input disturbance, which is addressed in the MPC design. This allows the neural network controller to directly inherit robustness from the underlying model predictive controller, which enables the transfer of guarantees to the approximation. The design process is an offline, constructive synthesis pipeline and can be implemented once Lipschitz bounds are specified. We further demonstrate in simulation that the learned controller reproduces the closed-loop behavior of the original controller while reducing online computation.
Hendrik Alsmeier, Maik Pfefferkorn, Rolf Findeisen
Control and Cyber-Physical Systems Laboratory, Technical University of Darmstadt, Landgraf-Georg-Str. 4, 64283 Darmstadt, Germany
We prove that the set NN of ReLU networks with one hidden layer and at most N∈N neurons is weakly closed in W1,p(Ω), for a bounded connected open set Ω⊂Rd. Consequently, every coercive, weakly lower semicontinuous energy on this space attains its minimum over NN. This covers the Deep Ritz method with convex integrands, Sobolev training and Lipschitz learning. Moreover, as the weight decay parameter tends to zero, minimizers of the regularized problem converge, up to subsequences, to minimizers with minimal parameter norm. We also prove that the set of ReLU networks with two hidden layers of bounded widths is not closed in W1,p(Ω) or in C(Ω). This disproves a conjecture of Petersen, Raslan and Voigtlaender.
Model Predictive Control (MPC) provides a structured and constraint-aware mechanism for decision-making, but its reliance on optimization-friendly analytical dynamics models limits its use in tasks with contacts and other hard-to-model state dependencies. Model-free reinforcement learning avoids explicit modeling assumptions but typically requires large amounts of interaction data. We present a learning-based MPC framework that combines the data efficiency and structure of local model-based planning with learned components that compensate for incomplete dynamics and finite-horizon myopia. The method augments a nominal analytical model with a residual dynamics network that learns missing state-dependent effects from data and combines the resulting planner with a learned action-value critic that injects long-horizon MDP structure into the local iLQR optimization. To make this practical at reinforcement-learning scale, we develop a GPU-accelerated batched iLQR solver that evaluates learned dynamics and critic networks inside the optimal-control loop and solves thousands of trajectory-optimization problems in parallel. The complete system is integrated into a robotics simulator, enabling scalable model-based reinforcement learning under incomplete dynamics. Experiments on biased and incompletely modeled control tasks show that the approach improves closed-loop control performance while preserving the model-based structure needed for efficient constrained trajectory optimization.
Philipp Schoch, Markus Ryll
Autonomous Aerial Systems Lab Technical University of Munich