Abstract
Two of the main challenges in optimal control are solving problems with state-dependent running costs and developing efficient numerical solvers that are computationally tractable in high dimensions. In this paper, we provide analytical solutions to certain optimal control problems whose running cost depends on the state variable and with constraints on the control. We also provide Lax-Oleinik-type representation formulas for the corresponding Hamilton-Jacobi partial differential equations with state-dependent Hamiltonians. Additionally, we present an efficient, grid-free numerical solver based on our representation formulas, which is shown to scale linearly with the state dimension, and thus, to overcome the curse of dimensionality. Using existing optimization methods and the min-plus technique, we extend our numerical solvers to address more general classes of convex and nonconvex initial costs. We demonstrate the capabilities of our numerical solvers using implementations on a central processing unit (CPU) and a field-programmable gate array (FPGA). In several cases, our FPGA implementation obtains over a 10 times speedup compared to the CPU, which demonstrates the promising performance boosts FPGAs can achieve. Our numerical results show that our solvers have the potential to serve as a building block for solving broader classes of high-dimensional optimal control problems in real-time.
Similar content being viewed by others
Data Availability
Data sharing is not applicable to this article as no new data were created or analyzed in this study.
References
Aĭpanov, S.A., Murzabekov, Z.N.: Analytical solution of a linear quadratic optimal control problem with control value constraints on the value of the control. J. Comput. Syst. Sci. Int. 53, 84–91 (2014). https://doi.org/10.1134/s1064230713060026
Akian, M., Bapat, R., Gaubert, S.: Max-plus algebra. In: Hogben, L. (ed) Handbook of Linear Algebra, vol. 39, pp. 10–14. Chapman and Hall/CRC, Boca Raton (2006)
Akian, M., Gaubert, S., Lakhoua, A.: The max-plus finite element method for solving deterministic optimal control problems: basic properties and convergence analysis. SIAM J. Control. Optim. 47(2), 817–848 (2008)
Alla, A., Falcone, M., Saluzzi, L.: An efficient DP algorithm on a tree-structure for finite horizon optimal control problems. SIAM J. Sci. Comput. 41(4), A2384–A2406 (2019)
Alla, A., Falcone, M., Volkwein, S.: Error analysis for POD approximations of infinite horizon problems via the dynamic programming approach. SIAM J. Control. Optim. 55(5), 3091–3115 (2017)
Bachouch, A., Huré, C., Langrené, N., Pham, H.: Deep neural networks algorithms for stochastic control problems on finite horizon: numerical applications. Methodol. Comput. Appl. Probab. 24(1), 143–178 (2022). https://doi.org/10.1007/s11009-019-09767-9
Bansal, S., Tomlin, C.: Deepreach: a deep learning approach to high-dimensional reachability. In: 2021 IEEE International Conference on Robotics and Automation (ICRA), Xi’an, China, 2021, pp. 1817–1824 (2021)
Bardi, M., Capuzzo-Dolcetta, I.: Optimal Control and Viscosity Solutions of Hamilton-Jacobi-Bellman Equations. Systems & Control: Foundations & Applications. Birkhäuser Boston, Inc., Boston (1997). https://doi.org/10.1007/978-0-8176-4755-1 (With appendices by Maurizio Falcone and Pierpaolo Soravia)
Bellman, R.E.: Adaptive Control Processes: a Guided Tour. Princeton University Press, Princeton (1961)
Bertsekas, D.P.: Reinforcement Learning and Optimal Control. Athena Scientific, Belmont (2019)
Bokanowski, O., Garcke, J., Griebel, M., Klompmaker, I.: An adaptive sparse grid semi-Lagrangian scheme for first order Hamilton-Jacobi Bellman equations. J. Sci. Comput. 55(3), 575–605 (2013)
Boyd, S., Parikh, N., Chu, E., Peleato, B., Eckstein, J.: Distributed optimization and statistical learning via the alternating direction method of multipliers. Found. Trends Mach. Learn. 3(1), 1–122 (2011). https://doi.org/10.1561/2200000016
Burachik, R.S., Kaya, C.Y., Majeed, S.N.: A duality approach for solving control-constrained linear-quadratic optimal control problems. SIAM J. Control. Optim. 52(3), 1423–1456 (2014). https://doi.org/10.1137/130910221
Cannon, M., Liao, W., Kouvaritakis, B.: Efficient MPC optimization using Pontryagin’s minimum principle. In: Proceedings of the 45th IEEE Conference on Decision and Control, pp. 5459–5464 (2006). https://doi.org/10.1109/CDC.2006.377753
Chen, J., Zhan, W., Tomizuka, M.: Constrained iterative LQR for on-road autonomous driving motion planning. In: 2017 IEEE 20th International Conference on Intelligent Transportation Systems (ITSC), pp. 1–7 (2017). https://doi.org/10.1109/ITSC.2017.8317745
Chen, J., Zhan, W., Tomizuka, M.: Autonomous driving motion planning with constrained iterative LQR. IEEE Trans. Intell. Veh. 4(2), 244–254 (2019). https://doi.org/10.1109/TIV.2019.2904385
Chen, M., Hu, Q., Fisac, J.F., Akametalu, K., Mackin, C., Tomlin, C.J.: Reachability-based safety and goal satisfaction of unmanned aerial platoons on air highways. J. Guid. Control. Dyn. 40(6), 1360–1373 (2017). https://doi.org/10.2514/1.G000774
Coupechoux, M., Darbon, J., Kèlif, J., Sigelle, M.: Optimal trajectories of a UAV base station using Lagrangian mechanics. In: IEEE INFOCOM 2019—IEEE Conference on Computer Communications Workshops (INFOCOM WKSHPS), pp. 626–631 (2019)
Darbon, J.: On convex finite-dimensional variational methods in imaging sciences and Hamilton-Jacobi equations. SIAM J. Imag. Sci. 8(4), 2268–2293 (2015). https://doi.org/10.1137/130944163
Darbon, J., Dower, P.M., Meng, T.: Neural network architectures using min-plus algebra for solving certain high-dimensional optimal control problems and Hamilton-Jacobi PDEs. Math. Control Signals Syst. 1–44 (2022)
Darbon, J., Langlois, G.P., Meng, T.: Overcoming the curse of dimensionality for some Hamilton-Jacobi partial differential equations via neural network architectures. Res. Math. Sci. 7(3), 20 (2020). https://doi.org/10.1007/s40687-020-00215-6
Darbon, J., Meng, T.: On decomposition models in imaging sciences and multi-time Hamilton-Jacobi partial differential equations. SIAM J. Imag. Sci. 13(2), 971–1014 (2020). https://doi.org/10.1137/19M1266332
Darbon, J., Meng, T.: On some neural network architectures that can represent viscosity solutions of certain high dimensional Hamilton-Jacobi partial differential equations. J. Comput. Phys. 425, 109907 (2021). https://doi.org/10.1016/j.jcp.2020.109907
Darbon, J., Meng, T., Resmerita, E.: On Hamilton-Jacobi PDEs and image denoising models with certain nonadditive noise. J. Math. Imaging Vis. 64(4), 408–441 (2022)
Darbon, J., Osher, S.: Algorithms for overcoming the curse of dimensionality for certain Hamilton-Jacobi equations arising in control theory and elsewhere. Res. Math. Sci. 3(19), 1–26 (2016). https://doi.org/10.1186/s40687-016-0068-7
Davis, D., Yin, W.: Faster convergence rates of relaxed Peaceman-Rachford and ADMM under regularity assumptions. Math. Oper. Res. 42(3), 783–805 (2017). https://doi.org/10.1287/moor.2016.0827
Delahaye, D., Puechmorel, S., Tsiotras, P., Feron, E.: Mathematical models for aircraft trajectory design: a survey. In: Air Traffic Management and Systems, pp. 205–247. Springer Japan, Tokyo (2014)
Deng, W., Yin, W.: On the global and linear convergence of the generalized alternating direction method of multipliers. J. Sci. Comput. 66(3), 889–916 (2016). https://doi.org/10.1007/s10915-015-0048-x
Denk, J., Schmidt, G.: Synthesis of a walking primitive database for a humanoid robot using optimal control techniques. In: Proceedings of IEEE-RAS International Conference on Humanoid Robots, pp. 319–326 (2001)
Djeridane, B., Lygeros, J.: Neural approximation of PDE solutions: an application to reachability computations. In: Proceedings of the 45th IEEE Conference on Decision and Control, pp. 3034–3039 (2006). https://doi.org/10.1109/CDC.2006.377184
Dolgov, S., Kalise, D., Kunisch, K.K.: Tensor decomposition methods for high-dimensional Hamilton-Jacobi-Bellman equations. SIAM J. Sci. Comput. 43(3), A1625–A1650 (2021). https://doi.org/10.1137/19M1305136
Dower, P.M., McEneaney, W.M., Cantoni, M.: Game representations for state constrained continuous time linear regulator problems. arXiv:1904.05552 (2019)
Dower, P.M., McEneaney, W.M., Zhang, H.: Max-plus fundamental solution semigroups for optimal control problems. In: 2015 Proceedings of the Conference on Control and Its Applications, pp. 368–375. SIAM (2015)
El Khoury, A., Lamiraux, F., Taïx, M.: Optimal motion planning for humanoid robots. In: 2013 IEEE International Conference on Robotics and Automation, pp. 3136–3141 (2013). https://doi.org/10.1109/ICRA.2013.6631013
Fallon, M., Kuindersma, S., Karumanchi, S., Antone, M., Schneider, T., Dai, H., D’Arpino, C.P., Deits, R., DiCicco, M., Fourie, D., Koolen, T., Marion, P., Posa, M., Valenzuela, A., Yu, K.-T., Shah, J., Iagnemma, K., Tedrake, R., Teller, S.: An architecture for online affordance-based perception and whole-body planning. J. Field Robot. 32(2), 229–254 (2015)
Feng, S., Whitman, E., Xinjilefu, X., Atkeson, C.G.: Optimization based full body control for the atlas robot. In: 2014 IEEE-RAS International Conference on Humanoid Robots, pp. 120–127 (2014). https://doi.org/10.1109/HUMANOIDS.2014.7041347
Fleming, W., McEneaney, W.: A max-plus-based algorithm for a Hamilton-Jacobi-Bellman equation of nonlinear filtering. SIAM J. Control. Optim. 38(3), 683–710 (2000). https://doi.org/10.1137/S0363012998332433
Fujiwara, K., Kajita, S., Harada, K., Kaneko, K., Morisawa, M., Kanehiro, F., Nakaoka, S., Hirukawa, H.: An optimal planning of falling motions of a humanoid robot. In: 2007 IEEE/RSJ International Conference on Intelligent Robots and Systems, pp. 456–462 (2007). https://doi.org/10.1109/IROS.2007.4399327
Garcke, J., Kröner, A.: Suboptimal feedback control of PDEs by solving HJB equations on adaptive sparse grids. J. Sci. Comput. 70(1), 1–28 (2017)
Gaubert, S., McEneaney, W., Qu, Z.: Curse of dimensionality reduction in max-plus based approximation methods: theoretical estimates and improved pruning algorithms. In: 2011 50th IEEE Conference on Decision and Control and European Control Conference, pp. 1054–1061. IEEE (2011)
Glowinski, R.: On alternating direction methods of multipliers: a historical perspective. In: Fitzgibbon, W., Kuznetsov, Y., Neittaanmäki, P., Pironneau, O. (eds) Modeling, Simulation and Optimization for Science and Technology. Computational Methods in Applied Sciences, vol. 34, pp. 59–82. Springer, Dordrecht (2014). https://doi.org/10.1007/978-94-017-9054-3_4
Han, J., Jentzen, A., E, W.: Solving high-dimensional partial differential equations using deep learning. Proc. Natl. Acad. Sci. 115(34), 8505–8510 (2018). https://doi.org/10.1073/pnas.1718942115
Hofer, M., Muehlebach, M., D’Andrea, R.: Application of an approximate model predictive control scheme on an unmanned aerial vehicle. In: 2016 IEEE International Conference on Robotics and Automation (ICRA), pp. 2952–2957 (2016). https://doi.org/10.1109/ICRA.2016.7487459
Horowitz, M.B., Damle, A., Burdick, J.W.: Linear Hamilton Jacobi Bellman equations in high dimensions. In: 53rd IEEE Conference on Decision and Control, pp. 5880–5887. IEEE (2014)
Hu, C., Shu, C.-W.: A discontinuous Galerkin finite element method for Hamilton-Jacobi equations. SIAM J. Sci. Comput. 21(2), 666–690 (1999). https://doi.org/10.1137/S1064827598337282
Huré, C., Pham, H., Bachouch, A., Langrené, N.: Deep neural networks algorithms for stochastic control problems on finite horizon: convergence analysis. SIAM J. Numer. Anal. 59(1), 525–557 (2021). https://doi.org/10.1137/20M1316640
Huré, C., Pham, H., Warin, X.: Deep backward schemes for high-dimensional nonlinear PDEs. Math. Comp. 89(324), 1547–1579 (2020). https://doi.org/10.1090/mcom/3514
Jaddu, H.: Spectral method for constrained linear-quadratic optimal control. Math. Comput. Simul. 58(2), 159–169 (2002). https://doi.org/10.1016/S0378-4754(01)00359-7
Jiang, F., Chou, G., Chen, M., Tomlin, C.J.: Using neural networks to compute approximate and guaranteed feasible Hamilton-Jacobi-Bellman PDE solutions. arXiv:1611.03158 (2016)
Jiang, G., Peng, D.: Weighted ENO schemes for Hamilton-Jacobi equations. SIAM J. Sci. Comput. 21(6), 2126–2143 (2000). https://doi.org/10.1137/S106482759732455X
Jin, L., Li, S., Yu, J., He, J.: Robot manipulator control using neural networks: a survey. Neurocomputing 285, 23–34 (2018). https://doi.org/10.1016/j.neucom.2018.01.002
Jin, P., Zhang, Z., Kevrekidis, I.G., Karniadakis, G.E.: Learning Poisson systems and trajectories of autonomous systems via Poisson neural networks. IEEE Trans. Neural Netw. Learn. Syst. (2022). https://doi.org/10.1109/TNNLS.2022.3148734
Jin, P., Zhang, Z., Zhu, A., Tang, Y., Karniadakis, G.E.: SympNets: intrinsic structure-preserving symplectic networks for identifying Hamiltonian systems. Neural Netw. 132, 166–179 (2020). https://doi.org/10.1016/j.neunet.2020.08.017
Kalise, D., Kundu, S., Kunisch, K.: Robust feedback control of nonlinear PDEs by numerical approximation of high-dimensional Hamilton-Jacobi-Isaacs equations. SIAM J. Appl. Dyn. Syst. 19(2), 1496–1524 (2020). https://doi.org/10.1137/19M1262139
Kalise, D., Kunisch, K.: Polynomial approximation of high-dimensional Hamilton-Jacobi-Bellman equations and applications to feedback control of semilinear parabolic PDEs. SIAM J. Sci. Comput. 40(2), A629–A652 (2018)
Kang, W., Wilcox, L.C.: Mitigating the curse of dimensionality: sparse grid characteristics method for optimal feedback control and HJB equations. Comput. Optim. Appl. 68(2), 289–315 (2017)
Kastner, R., Matai, J., Neuendorffer, S.: Parallel Programming for FPGAs. arXiv:1805.03648v1 (2018)
Kim, Y.H., Lewis, F.L., Dawson, D.M.: Intelligent optimal control of robotic manipulators using neural networks. Automatica 36(9), 1355–1364 (2000). https://doi.org/10.1016/S0005-1098(00)00045-5
Kolokoltsov, V.N., Maslov, V.P.: Idempotent Analysis and Its Applications. Mathematics and Its Applications, vol. 401. Kluwer Academic Publishers Group, Dordrecht (1997). https://doi.org/10.1007/978-94-015-8901-7 (Translation of ıt Idempotent analysis and its application in optimal control (Russian), “Nauka” Moscow, 1994 [ MR1375021 (97d:49031)], Translated by V. E. Nazaikinskii, With an appendix by Pierre Del Moral)
Kuindersma, S., Deits, R., Fallon, M., Valenzuela, A., Dai, H., Permenter, F., Koolen, T., Marion, P., Tedrake, R.: Optimization-based locomotion planning, estimation, and control design for the atlas humanoid robot. Auton. Robot. 40(3), 429–455 (2016)
Kunisch, K., Volkwein, S., Xie, L.: HJB-POD-based feedback design for the optimal control of evolution problems. SIAM J. Appl. Dyn. Syst. 3(4), 701–722 (2004)
Lambrianides, P., Gong, Q., Venturi, D.: A new scalable algorithm for computational optimal control under uncertainty. J. Comput. Phys. 420, 109710 (2020). https://doi.org/10.1016/j.jcp.2020.109710
Lee, D., Tomlin, C.J.: A computationally efficient Hamilton-Jacobi-based formula for state-constrained optimal control problems. arXiv:2106.13440 (2021)
Lee, D., Tomlin, C.J.: A Hopf-Lax formula in Hamilton-Jacobi analysis of reach-avoid problems. IEEE Control Syst. Lett. 5(3), 1055–1060 (2021). https://doi.org/10.1109/LCSYS.2020.3009933
Lewis, F., Dawson, D., Abdallah, C.: Robot Manipulator Control: Theory and Practice. Control Engineering. Marcel Dekker Inc., New York (2004). https://books.google.com/books?id=BDS_PQAACAAJ
Li, A., Bansal, S., Giovanis, G., Tolani, V., Tomlin, C., Chen, M.: Generating robust supervision for learning-based visual navigation using Hamilton-Jacobi reachability. In: Bayen, A.M., Jadbabaie, A., Pappas, G., Parrilo, P.A., Recht, B., Tomlin, C., Zeilinger, M. (eds.) Proceedings of the 2nd Conference on Learning for Dynamics and Control, Proceedings of Machine Learning Research, vol. 120, pp. 500–510. PMLR, The Cloud (2020). http://proceedings.mlr.press/v120/li20a.html
Li, W., Todorov, E.: Iterative linear quadratic regulator design for nonlinear biological movement systems. In: 2004 International Conference on Informatics in Control, Automation and Robotics, pp. 222–229. Citeseer (2004)
Lin, F., Brandt, R.D.: An optimal control approach to robust control of robot manipulators. IEEE Trans. Robot. Automat. 14(1), 69–77 (1998). https://doi.org/10.1109/70.660845
Ma, J., Cheng, Z., Zhang, X., Tomizuka, M., Lee, T.H.: Alternating direction method of multipliers for constrained iterative LQR in autonomous driving. IEEE Trans. Intell. Transp. Syst. 23, 23031–23042 (2022). https://doi.org/10.1109/TITS.2022.3194571
McEneaney, W.: A curse-of-dimensionality-free numerical method for solution of certain HJB PDEs. SIAM J. Control. Optim. 46(4), 1239–1276 (2007). https://doi.org/10.1137/040610830
McEneaney, W.M.: Max-Plus Methods for Nonlinear Control and Estimation. Systems & Control: Foundations & Applications. Birkhäuser Boston, Inc., Boston (2006)
McEneaney, W.M., Deshpande, A., Gaubert, S.: Curse-of-complexity attenuation in the curse-of-dimensionality-free method for HJB PDEs. In: 2008 American Control Conference, pp. 4684–4690. IEEE (2008)
McEneaney, W.M., Dower, P.M.: The principle of least action and fundamental solutions of mass-spring and \(N\)-body two-point boundary value problems. SIAM J. Control. Optim. 53(5), 2898–2933 (2015)
McEneaney, W.M., Kluberg, L.J.: Convergence rate for a curse-of-dimensionality-free method for a class of HJB PDEs. SIAM J. Control. Optim. 48(5), 3052–3079 (2009)
Nakamura-Zimmerer, T., Gong, Q., Kang, W.: Adaptive deep learning for high-dimensional Hamilton-Jacobi-Bellman equations. SIAM J. Sci. Comput. 43(2), A1221–A1247 (2021). https://doi.org/10.1137/19M1288802
Nakamura-Zimmerer, T., Gong, Q., Kang, W.: QRnet: optimal regulator design with LQR-augmented neural networks. IEEE Control Syst. Lett. 5(4), 1303–1308 (2021). https://doi.org/10.1109/LCSYS.2020.3034415
Niarchos, K.N., Lygeros, J.: A neural approximation to continuous time reachability computations. In: Proceedings of the 45th IEEE Conference on Decision and Control, pp. 6313–6318 (2006). https://doi.org/10.1109/CDC.2006.377358
Onken, D., Nurbekyan, L., Li, X., Fung, S.W., Osher, S., Ruthotto, L.: A neural network approach for high-dimensional optimal control applied to multiagent path finding. IEEE Trans. Control Syst. Technol. (2022). https://doi.org/10.1109/TCST.2022.3172872
Osher, S., Shu, C.-W.: High-order essentially nonoscillatory schemes for Hamilton-Jacobi equations. SIAM J. Numer. Anal. 28(4), 907–922 (1991). https://doi.org/10.1137/0728049
Park, J.H., Han, S., Kwon, W.H.: LQ tracking controls with fixed terminal states and their application to receding horizon controls. Syst. Control Lett. 57(9), 772–777 (2008). https://doi.org/10.1016/j.sysconle.2008.03.006
Parzani, C., Puechmorel, S.: On a Hamilton-Jacobi-Bellman approach for coordinated optimal aircraft trajectories planning. In: CCC 2017 36th Chinese Control Conference (CCC), Dalian, China, pp. 353–358. IEEE (2017). https://doi.org/10.23919/ChiCC.2017.8027369. https://hal-enac.archives-ouvertes.fr/hal-01340565
Prakash, S.K.: Managing HBM’s Bandwidth in Multi-die FPGAs Using Overlay NoCs. Master’s thesis, University of Waterloo (2021)
Reisinger, C., Zhang, Y.: Rectified deep neural networks overcome the curse of dimensionality for non-smooth value functions in zero-sum games of nonlinear stiff systems. Anal. Appl. (Singap.) 18(6), 951–999 (2020). https://doi.org/10.1142/S0219530520500116
Rockafellar, R.T., Wets, R.J.B.: Variational Analysis, Grundlehren der Mathematischen Wissenschaften [Fundamental Principles of Mathematical Sciences], vol. 317. Springer, Berlin (1998). https://doi.org/10.1007/978-3-642-02431-3
Royo, V.R., Tomlin, C.: Recursive regression with neural networks: approximating the HJI PDE solution. arXiv:1611.02739 (2016)
Rucco, A., Sujit, P.B., Aguiar, A.P., de Sousa, J.B., Pereira, F.L.: Optimal rendezvous trajectory for unmanned aerial-ground vehicles. IEEE Trans. Aerosp. Electron. Syst. 54(2), 834–847 (2018). https://doi.org/10.1109/TAES.2017.2767958
Russo, D.: Adaptation of High Performance and High Capacity Reconfigurable Systems to OpenCL Programming Environments. Master’s thesis, Universitat Politècnica de València (2020)
Sideris, A., Bobrow, J.E.: An efficient sequential linear quadratic algorithm for solving nonlinear optimal control problems. In: Proceedings of the 2005, American Control Conference, vol. 4, pp. 2275–2280. IEEE (2005). https://doi.org/10.1109/ACC.2005.1470308
Sirignano, J., Spiliopoulos, K.: DGM: a deep learning algorithm for solving partial differential equations. J. Comput. Phys. 375, 1339–1364 (2018). https://doi.org/10.1016/j.jcp.2018.08.029
Todorov, E.: Efficient computation of optimal actions. Proc. Natl. Acad. Sci. 106(28), 11478–11483 (2009)
Yegorov, I., Dower, P.M.: Perspectives on characteristics based curse-of-dimensionality-free numerical approaches for solving Hamilton-Jacobi equations. Appl. Math. Optim. 83, 1–49 (2021)
Zhang, H., Dower, P.M.: A max-plus based fundamental solution for a class of discrete time linear regulator problems. Linear Algebra Appl. 471, 693–729 (2015)
Zhou, M., Han, J., Lu, J.: Actor-critic method for high dimensional static Hamilton-Jacobi-Bellman partial differential equations based on neural networks. SIAM J. Sci. Comput. 43(6), A4043–A4066 (2021). https://doi.org/10.1137/21M1402303
Acknowledgements
This research is supported by the DOE-MMICS SEA-CROGS DE-SC0023191 and the AFOSR MURI FA9550-20-1-0358. P.C. is supported by the SMART Scholarship, which is funded by the USD/R &E (The Under Secretary of Defense-Research and Engineering), National Defense Education Program (NDEP) / BA-1, Basic Research. We thank Peter Dower for his useful feedback.
Author information
Authors and Affiliations
Corresponding author
Ethics declarations
Conflict of Interest
The authors declare that they have no known competing financial interests or personal relationships that could have influenced or appeared to have influenced the work reported in this paper. Furthermore, the authors declare that they have no known conflicts of interest.
Additional information
Authors’ names are given in last/family name alphabetical order.
Appendices
Appendix A Some Technical Lemmas for the Analytical Solutions
Lemma A1
Let a, b, t be positive scalars and \(x,u\) be real numbers satisfying \(u-bt\leqslant x\leqslant u+at\). Let \(V\) be the function defined in (9) and (15) and \([0,t]\ni s\mapsto \gamma (s;x,t,u,a,b)\in \mathbb {R}\) be the trajectory defined in (12), (13), (14), and (16) for different cases. Then, there holds
Proof
If \((x,t,u)\in \varOmega _1\) holds, we have
If \((x,t,u)\in \varOmega _2\) holds, we have
If \((x,t,u)\in \varOmega _3\) holds, we have
If \(u<0\), we have
Therefore, (A1) holds for any \((x,t,u)\in \mathbb {R}\times (0,+\infty )\times \mathbb {R}\) satisfying \(u-bt\leqslant x\leqslant u+at\).
Lemma A2
Let a, b be positive scalars. Let \(V\) be the function defined in (9) and (15). Then, for any \(x\in \mathbb {R}\), \(t>0\), the function \(\mathbb {R}\ni u\mapsto V(x,t;u,a,b)\in \mathbb {R}\cup \{+\infty \}\) is strictly convex and twice continuously differentiable in its domain.
Proof
In this proof, we regard the function \(V(x,t;u,a,b)\) as a function of \(u\) from its domain \([x-at,x+bt]\) to \(\mathbb {R}\), and we use its derivative to mean the derivative of \(V\) with respect to \(u\), by default. To prove the statement, we need to prove that \(V\) is twice continuously differentiable and that the second-order derivative is positive almost everywhere in the domain. We consider the following cases.
First, assume \(x\geqslant at\) holds. After some computation, the function \(u\mapsto V(x,t;u,a,b)\) can be written as follows:
which is twice continuously differentiable. The second-order derivative is given by
where the first and second inequalities hold since we have \(u\geqslant x-at\geqslant 0\). Moreover, the second inequality becomes equality if and only if \(u\) is zero. In other words, the second-order derivative in (A2) is positive almost everywhere, and hence, the conclusion holds in this case.
Next, assume that x is a point in [0, at). In this case, the function \(u\mapsto V(x,t;u,a,b)\) can be written as follows:
It is straightforward to check that this function is twice continuously differentiable and that the second-order derivative reads
where the first line is positive since \(u<0\) holds in the first line, the second line is positive almost everywhere since \(u>0\) holds almost everywhere in the second line, and the third line is positive since the inequalities in (A2) also hold according to the condition on u (there holds \(u \geqslant bt-\frac{bx}{a}>0>x-at\)). Therefore, the conclusion follows in this case.
Finally, we consider the case when \(x<0\). By definition, we have that \(V(x,t;u,a,b) = V(-x,t;-u,b,a)\), where the right-hand side is twice continuously differentiable and whose second-order derivative with respect to \(-u\) is positive almost everywhere by the same argument above. Therefore, the function \(V(x,t;u,a,b)\) is also strictly convex and twice continuously differentiable with respect to \(u\), and the conclusion holds.
Appendix B Some Computations for the Numerical Implementation
1.1 Appendix B.1 A Numerical Method for Computing the Proximal Point of \(u\mapsto \frac{1}{\lambda }V(x,t;u,a,b)\)
Here, we discuss how to compute the proximal point of the function \(\mathbb {R}\ni u\mapsto \frac{1}{\lambda }V(x,t;u,a,b)\in \mathbb {R}\cup \{+\infty \}\), i.e., how to solve the following convex optimization problem:
for any \(\lambda ,t,a,b>0\), and \(x,y\in \mathbb {R}\). We consider the following two cases for the variable u.
If \(u \leqslant 0\), after some computation, the objective function in (B1) can be written as
where the three regions \(\varOmega _1(x,t,a,b), \varOmega _2(x,t,a,b), and~\varOmega _3(x,t,a,b)\subset [0,+\infty )\) are defined by
In this case, the derivative of F with respect to u is given by
and the second derivative of F with respect to u can be easily computed using (A2) and (A3), for different cases. To get possible candidates for the minimizer \(u^*\) of F in this case, we compute the roots of the functions in the two lines of (B2) and select the roots where the second derivative of F is non-negative. After some calculations, the candidates are given by \(u_1\) and \(u_2\), which are defined as follows:
Note that \(u_1\) and \(u_2\) may be not well-defined if the term under the square root is negative, in which case the corresponding function does not provide a possible candidate for \(u^*\). Therefore, we assign \(u_i\) (\(i=1,2\)) to be an arbitrary point in \([x-at,x+bt]\) if it is not well-defined.
If \(u < 0\), after some computation, the objective function in (B1) can be written as
Thus, for \(u < 0\), the derivative of F with respect to u is given by
Similarly as in the first case, we take the roots of the two functions in (B4), such that the second order derivative of F is non-negative. These roots provide possible candidates for \(u^*\). We denote these candidates by \(u_1'\) and \(u_2'\), which are defined by
Similarly, if \(u_1'\) or \(u_2'\) is not well-defined, we set it to be any point in \([x-at,x+bt]\).
Note that the objective function F is strictly convex and twice continuously differentiable with respect to u by Lemma A2. Then, by the first and second derivative tests, the minimizer \(u^*\) in (B1) is selected among the possible candidates \(u_1,u_2,u'_1, u_2'\) defined in (B3) and (B5), as well as the boundary points \(x-at\) and \(x+bt\). In other words, the minimizer \(u^*\) satisfies
Numerically, we solve the optimization problem (B1) by computing the six candidates \(u_1,u_2,u_1',u_2',x-at,x+bt\) and comparing the objective function values at those points. Then, the minimizer \(u^*\) is selected using (B6). Therefore, the complexity of solving (B1) is \(\varTheta (1)\).
1.2 Appendix B.2 An Equivalent Expression for \(V(x,t;u,a,b)\) and \(\gamma (s;x,t,u,a,b)\)
Let \(V\) be the function defined by (9) and (15), and let \(\gamma \) be the function defined by (12), (13), (14), and (16) for different cases. Now, we present an equivalent expression for \(V\) and \(\gamma \), which is used in our numerical implementation.
By straightforward calculation, the function \(V\) can equivalently be expressed as
where \(V_1,V_2\), and \(V_3\) are the functions in the first, second, and third lines of (9), respectively.
Similarly, assuming \(u\in [x-at,x+bt]\) holds, the function \(\gamma \) can be expressed as
Compared to the definitions (9), (15), (12), (13), (14), and (16), these equivalent formulas involve less conditional branching, and hence, they are more favorable for the performance of our numerical implementation.
Appendix C Proofs of Convergence Results in Sect. 3
In this Appendix, we provide the proof of Proposition 7 in Appendix C.1 and the proof of Proposition 8 in Appendix C.2.
1.1 Appendix C.1 Proof of Proposition 7
Let \(\varvec{v}^N\), \(\varvec{d}^N\), and \(\varvec{u}^N\) be the corresponding vectors in the algorithm at the N-th iteration. Let \(\varvec{u}^*\) be the minimizer of the minimization problem in (24), which is unique since \(\varPhi \) is convex and each \(u_i\mapsto V(x_i,t;u_i,a,b)\) is strictly convex by Lemma A2. According to [28, Theorem 2.2] whose assumptions are proved using [28, Remark 2.2], both \(\varvec{v}^N\) and \(\varvec{d}^N\) converge to the point \(\varvec{u}^*\) as N approaches infinity, and hence, \(\varvec{u}^N\) in Algorithm 1 also converges to \(\varvec{u}^*\). Since \(\varPhi \) is a real-valued convex function, it is continuous in \(\mathbb {R}^n\) and we have
Note that the domain of the function \(\varvec{u}\mapsto \sum _{i=1}^nV(x_i,t;u_i, a_i,b_i)\) equals
and the point \(\varvec{u}^N = \varvec{d}^N\) is in the set in (C2) by definition of \(\varvec{d}^N\). Thus, the point \(\varvec{u}^N\) is in the domain of the function \(\varvec{u}\mapsto \sum _{i=1}^nV(x_i,t;u_i, a_i,b_i)\). By Lemma A2, the function \(\varvec{u}\mapsto \sum _{i=1}^nV(x_i,t;u_i, a_i,b_i)\) is continuous in its domain. It is also straightforward to check that the function \(\varvec{u}\mapsto \sum _{i=1}^nV(x_i,t;u_i, a_i,b_i)\) is Lipschitz in its domain, and we denote its Lipschitz constant by \(L_i\). Thus, we have that
Then, the convergence of \(\hat{V}^N(\varvec{x},t)\) to \(V(\varvec{x},t)\) follows from (C1) and (C3).
Now, it remains to prove the second formula in (44). We have proved that \(\varvec{u}^N\) and \(\varvec{u}^*\) are both in the set in (C2). Let \(\{\varvec{u}^{N_j}\}_j\) be a subsequence of \(\{\varvec{u}^{N}\}_N\) (i.e., we assume \(N_1<N_2<\cdots \) and \(\lim _{j\rightarrow \infty }N_j=+\infty \)), such that for each \(i\in \{1,\cdots , n\}\), the i-th component \(\{u_i^{N_j}\}_j\) of the subsequence satisfies one of the following assumptions:
-
(i)
there exists an index \(r_i\) in \(\{1,2,3\}\), such that there hold
$$\begin{aligned} u_i^{N_j}\geqslant 0 \quad \text { and }\quad (x_i,t,u_i^{N_j})\in \bar{\varOmega }_{r_i}(a_i,b_i),\quad \forall \, j\in \mathbb {N}; \end{aligned}$$ -
(ii)
there exists an index \(r_i\) in \(\{1,2,3\}\), such that there hold
$$\begin{aligned} u_i^{N_j}< 0 \quad \text { and }\quad (-x_i,t,-u_i^{N_j})\in \bar{\varOmega }_{r_i}(b_i,a_i),\quad \forall \, j\in \mathbb {N}. \end{aligned}$$
Here, to emphasize the dependence on \(a_i\) and \(b_i\), we use \(\varOmega _{r_i}(a_i,b_i)\) and \(\bar{\varOmega }_{r_i}(a_i,b_i)\) to respectively denote the set defined in (10) with constants \(a=a_i\) and \(b=b_i\) and its closure. Note that the situations considered in cases (i) and (ii) give a partition of the set in (C2) (where some sets in the partition may be empty and the sets may overlap on the boundary, but neither of these possibilities affect the result). Hence, if the statement is proved for any such subsequence \(\{\varvec{u}^{N_j}\}_j\), then the statement also holds for the whole sequence \(\{\varvec{u}^{N}\}_N\). Thus, it suffices to prove the statement for the subsequence \(\{\varvec{u}^{N_j}\}_j\).
Let \(i\in \{1,\cdots ,n\}\) be any index. Assume case (i) holds for \(\{u_i^{N_j}\}_j\) with the index \(r_i\). In other words, we assume \((x_i,t,u_i^{N_j})\in \bar{\varOmega }_{r_i}(a_i,b_i)\) holds for any \(j\in \mathbb {N}\). Since the set \(\bar{\varOmega }_{r_i}(a_i,b_i)\) is closed and the subsequence \(\{u_i^{N_j}\}_j\) converges to \(u_i^*\), we conclude that \((x_i,t,u_i^*)\in \bar{\varOmega }_{r_i}(a_i,b_i)\) also holds. Then, by definition of \(\gamma \) in \(\varOmega _{r_i}(a_i,b_i)\), it is straightforward to check that
The proof for case (ii) is similar, so we omit it. Note that (C4) holds for any arbitrary index \(i\in \{1,\cdots ,n\}\). Hence, we have
where the second inequality holds by (C4). Thus, the second formula in (44) holds for the subsequence by the convergence of \(\varvec{u}^{N_j}\) to \(\varvec{u}^*\). Moreover, the argument holds for any such subsequence, and hence, the statement holds for the whole sequence.
1.2 Appendix C.2 Proof of Proposition 8
Let r be the index defined in (51), and let the index set \({\mathcal {J}}\subseteq \{1,\cdots ,m\}\) be defined by
Then, we have
where the first equality holds by definition of r and the first inequality holds since \(V\) satisfies (47). Similarly, for any \(j\in {\mathcal {J}}\), we have
where the first equality holds by definition of \({\mathcal {J}}\) and the first inequality holds since \(\hat{V}\) satisfies (52). Therefore, (54) holds.
Now, assume \(V_j(\varvec{x},t) > V(\varvec{x},t) + 2\epsilon \) holds for each index j satisfying \(V_j(\varvec{x},t) \ne V(\varvec{x},t)\). We prove \(r\in {\mathcal {J}}\) by contradiction. Assume r is not in \({\mathcal {J}}\). Then, we have \(V_r(\varvec{x},t) \ne V(\varvec{x},t)\). However, from straightforward calculation, we also have
which leads to a contradiction with our assumption. Therefore, we have \(r\in {\mathcal {J}}\), and hence (55) holds by definition of r and \({\mathcal {J}}\).
Rights and permissions
Springer Nature or its licensor (e.g. a society or other partner) holds exclusive rights to this article under a publishing agreement with the author(s) or other rightsholder(s); author self-archiving of the accepted manuscript version of this article is solely governed by the terms of such publishing agreement and applicable law.
About this article
Cite this article
Chen, P., Darbon, J. & Meng, T. Lax-Oleinik-Type Formulas and Efficient Algorithms for Certain High-Dimensional Optimal Control Problems. Commun. Appl. Math. Comput. (2024). https://doi.org/10.1007/s42967-024-00371-4
Received:
Revised:
Accepted:
Published:
DOI: https://doi.org/10.1007/s42967-024-00371-4
Keywords
- Optimal control
- Hamilton-Jacobi partial differential equations
- Grid-free numerical methods
- High dimensions
- Field-programmable gate arrays (FPGAs)