Successive convexification methods for model predictive control of nonlinear systems with continuous and discrete variables

The sequential convexization method reformulates MINLP problems into separable form using MICP subproblems, addressing the computational challenges of hybrid systems with both continuous and discrete variables, enabling efficient real-time solutions.

JP2024541698AActive Publication Date: 2024-11-08MITSUBISHI ELECTRIC CORP
View PDF 4 Cites 0 Cited by

Patent Information

Application Number
JP2024548798
Authority / Receiving Office
JP · JP
Patent Type
Applications
Current Assignee / Owner
Priority Date
2021-11-18
Filing Date
2022-09-01
Publication Date
2024-11-08
Estimated Expiration
2042-09-01

AI Technical Summary

Technical Problem

Existing methods for model predictive control of hybrid systems with both continuous and discrete variables are computationally difficult due to the non-convex nature of mixed integer nonlinear programming problems, making real-time solutions impractical.

Method used

A sequential convexization method that reformulates MINLP problems into separable form, using mixed integer convex programming (MICP) subproblems to iteratively update solutions and reduce computational complexity through partial convexity techniques and branch-and-bound methods.

Benefits of technology

Enables feasible and sometimes suboptimal but computationally efficient solutions for MINMPC problems, allowing real-time implementation in hybrid systems with both continuous and discrete variables.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure 2024541698000001_ABST
    Figure 2024541698000001_ABST
Patent Text Reader

Abstract

To control the hybrid dynamic system, the predictive feedback controller creates a mixed integer nonlinear programming (MINLP) problem that includes nonlinear functions of continuous optimization variables that represent continuous elements of the hybrid dynamic system's behavior and / or one or more linear functions of integer optimization variables that represent discrete elements of the hybrid dynamic system's behavior. The MINLP problem is formulated into a separable format that ensures that the discrete elements of the behavior exist only in linear functions of the MINLP problem. The MINLP problem is solved over multiple iterations using partial convexification of a portion of the solution space that includes the current solution guess. The partial convexification produces a convex approximation of the nonlinear function in the MINLP without approximating the linear function in the MINLP to produce the partially convexified MINLP.
Need to check novelty before this filing date? Find Prior Art

Description

[Technical field]

[0001] The present disclosure relates generally to mixed integer nonlinear optimization-based control, and more particularly to a successive convexification method and apparatus for model predictive control of systems described by nonlinear dynamics with continuous and discrete elements of operation. [Background technology]

[0002] Optimization-based decision-making, planning, and control techniques such as model predictive control (MPC) enable a model-based design framework in which system dynamics, system requirements, and constraints can be directly considered. This framework has been extended to hybrid dynamical systems that include both continuous and discrete decision variables, which provides a powerful technique for modeling a wide range of problems, including, for example, dynamical systems with mode switching or systems with quantized actuation, problems with logic rules, time-logic specifications, or obstacle avoidance constraints. However, the resulting optimization problems are highly non-convex and therefore difficult to solve in practice because they involve variables that can only take on discrete values ​​(e.g., binary or integer values). When using nonlinear system dynamics, one or more nonlinear constraint functions, and / or a nonlinear objective function, the resulting optimal control problem (OCP) can be formulated as a mixed-integer nonlinear programming (MINLP) problem, which is NP-hard and therefore computationally difficult to solve.

[0003] In general, mixed integer nonlinear model predictive control (MINMPC) requires the solution of a nonconvex MINLP, i.e., the optimization problem is nonconvex even after relaxing the integer constraints due to, for example, nonlinear system dynamics, nonlinear constraint functions, and / or objective functions. Most successful global optimization algorithms for MINLP require convexity of the objective and constraint functions and therefore cannot be used to solve MINMPC problems to global optimum. Although global optimization algorithms exist for nonconvex MINLP, for example using relaxation of factorizable problems, they are usually computationally very expensive and therefore generally not yet practical for real-time implementation of MINMPC.

[0004] Decision-making, planning, or control for hybrid systems aims to solve a MINLP at every sampling instant to enable real-time MINMPC applications. Thus, several methods focus on approximation or heuristic techniques to find feasible but (possibly) suboptimal solutions to MINLPs within strict timing requirements. Some existing techniques are based on global algorithms for convex MINLPs, which can be used to find approximate solutions to non-convex MINLPs, for example, using outer approximation or hybrid branch-and-bound (hB&B) methods. In particular for non-convex MINMPCs, variants of real-time iterative (RTI) algorithms based on outer convexification combined with rounding schemes have been proposed. However, the latter approach requires solving mathematical programs with vanishing constraints, which are particularly challenging when inequality constraints depend directly on discrete decision variables.

[0005] Sequential convexification techniques, for example using sequential convex programming (SCP) or sequential quadratic programming (SQP) methods, form a popular technique for solving general nonlinear programming (NLP) problems. In particular, sequential convexification techniques have been successfully used for real-time realization of nonlinear model predictive control (NMPC) with smooth nonlinear dynamics, nonlinear constraint functions and / or inequality constraints. However, for systems with both continuous elements of operation and discrete elements of operation, i.e., systems that include continuous as well as integer and / or binary decision variables in MINMPC, it is necessary to extend these methods to NMPC.

[0006] Recent prior work has proposed a mixed integer sequential quadratic programming (MISQP) method based on the use of trust region radii for both continuous and integer optimization variables. This method requires the solution of a mixed integer quadratic programming (MIQP) subproblem at each iteration, which can be solved efficiently using, for example, state-of-the-art branch-and-bound (B&B) optimization methods. However, standard MISQP methods rely on the assumption that integer variables have a smooth effect on MINLPs, i.e., incrementing an integer variable by 1 results in a small change in the function value. However, the latter assumption is generally not true for MINMPCs, because, for example, constrained optimization problems may contain binary variables that have a large effect on the optimal control trajectory.

[0007] There is therefore a need for a successive convexification method that is more generally applicable to MINMPC problems, which is the objective of the systems and methods described in this invention. Summary of the Invention

[0008] The embodiment of the present invention is based on the solution of a sequence of one or more Mixed Integer Convex Programming (MICP) subproblems, where the preparation of each subproblem is performed based on a partial convexification technique to calculate a feasible, but possibly suboptimal, solution of the mixed integer nonlinear optimal control problem at each sampling instant of the proposed MINMPC controller. The solution of each MICP subproblem can be used to calculate updates to optimal solution guesses for all integer and / or binary decision variables, as well as new search directions for the continuous decision variables. In addition, based on the updated values ​​for the integer and / or binary decision variables and based on the new search directions for the continuous decision variables, the current solution guesses for the continuous decision variables can be updated at each iteration of the Mixed Integer Sequential Convex Programming (MISCP) optimization algorithm.

[0009] Some embodiments of the invention are based on the realization that any MINLP can be reformulated in a separable form as a different but mathematically equivalent MINLP, where all integer and / or binary decision variables enter linearly in all constraint and objective functions. The latter reformulation can be achieved, for example, by defining one or more auxiliary continuous optimization variables to ensure that all integer and / or binary optimization variables enter linearly in the constraint and objective functions. In particular, all nonlinear functions, which may therefore be present in a separable form in the constraint and objective functions of the MINMPC formulation, depend only on continuous optimization variables. Some embodiments of the invention are based on the realization that the latter linear dependence of the constraint and objective functions in the MINMPC formulation on all integer and / or binary decision variables can be used to avoid the smoothness requirement, i.e., incrementing an integer variable by one results in a small change in the function value, which limits the applicability of standard MISQP algorithms to MINMPC.

[0010] A partial convexification technique is used to prepare the MICP subproblems in each iteration of the proposed MISCP optimization algorithm. Some embodiments of the present invention are based on the recognition that due to the linear dependence of the constraint and objective functions in the MINMPC formulation on all integer and / or binary decision variables, the partial convexification technique needs to be applied only to smooth nonlinear functions. In some embodiments of the present invention, the partial convexification technique is based on a local linearization of all smooth nonlinear functions that may be present in the constraint and objective functions of the MINMPC formulation based on a solution guess of the continuous optimization variables in the current iteration of the MISCP algorithm. Alternatively, in other embodiments of the present invention, the partial convexification technique can be used to calculate a more general convex approximation of the nonlinear functions in one or more inequality constraints of the MINMPC problem formulation, resulting in convex quadratic inequality constraints in a mixed-integer quadratic constrained quadratic programming (MIQCQP) subproblem, or convex quadratic cone constraints in a mixed-integer quadratic cone programming (MISOCP) subproblem.

[0011] The present invention is based on the recognition that the solutions of the MICP subproblems in each iteration of the proposed MISCP algorithm can be computed relatively quickly thanks to the progress made over the past decades in the development of state-of-the-art solvers for MICP. For example, branch-and-bound methods can be used to efficiently solve mixed-integer quadratic programming (MIQP) or mixed-integer linear programming (MILP) subproblems. State-of-the-art branch-and-bound methods for MIQP and / or MILP may include advanced primal heuristics, branching strategies, pre-solving processes, cut generation techniques, and convex solvers with early termination and infeasibility detection to effectively reduce the size of the branch-and-bound search tree and thus the amount of convex relaxations that need to be solved to compute a globally optimal solution for each MIQP / MILP.

[0012] Some embodiments of the present invention are based on the realization that the updates of the continuous decision variables in each MISCP iteration can be performed in a specific way to guarantee some progress in computing a feasible but possibly suboptimal solution of the MINLP problem. Some embodiments of the present invention use a globalization strategy based on a merit function that quantifies a combination of optimality and constraint satisfaction for solution guesses for values ​​of the continuous and discrete decision variables. One example of a merit function is based on an l1 penalty function applied to each of the equality constraints and each violation of the inequality constraints in the MINMPC problem formulation, except for the integer constraints, which are automatically satisfied by design for each of the integer and / or binary decision variables in each iteration of the proposed MISCP optimization algorithm.

[0013]

number

[0014] In some embodiments of the invention, the globalization strategy is based on a trust region method that computes a small region in the space of continuous decision variables where the local convex approximation of the nonlinear constraints and / or objective function is sufficiently accurate. The accuracy of the partial convexification can be approximately evaluated based on the ratio of the actual versus predicted reduction in the value of the merit function from one solution guess to the next in each iteration of the MISCP optimization algorithm. Some embodiments of the invention are based on a trust region radius, whose value can be increased, decreased, or kept the same in each iteration of the algorithm.

[0015] Some embodiments of the present invention are based on the recognition that a relatively small trust region radius may be used when the MISCP optimization algorithm is sufficiently close to a feasible, but possibly suboptimal, solution of the MINMPC problem, resulting in a significant reduction in the computational cost for solving the MICP subproblem due to the relatively small trust region radius.

[0016] Some embodiments of the present invention are based on a warm-start strategy that computes guesses for optimal values ​​of continuous and integer and / or binary decision variables based on a feasible but possibly suboptimal solution of the MINMPC problem at a previous sampling time instant. For example, some embodiments of the present invention may use a time-shift procedure by one sampling period to warm-start a solution guess for the MISCP optimization algorithm given an optimal or near-optimal solution to the MINLP at a previous sampling time instant of the proposed MINMPC controller.

[0017] Some embodiments of the present invention are based on the solution of a limited number of one or more MICP subproblems, after which some or all of the integer and / or binary decision variables remain fixed, resulting in some one or more solutions of a convex programming (CP) subproblem. An upper bound on the number of MICP solutions can be chosen to be relatively small so that the computational cost of the MISCP optimization algorithm can be significantly reduced, while being sufficiently large to allow the MISCP optimization algorithm to compute well-feasible, but possibly suboptimal, solutions to the MINMPC problem. For example, in some embodiments of the present invention, only a single MICP subproblem needs to be solved at each sampling instant of a real-time executable MINMPC controller.

[0018] Some embodiments of the present invention are based on the recognition that the solution of one or more CP subproblems can be performed at significantly less computational cost than the solution of one or more MICP subproblems. Some embodiments of the present invention are based on the recognition that fixing some or all of the integer and / or binary decision variables after the solution of one or more of a limited number of MICP subproblems can prevent cycles of the MISCP optimization algorithm in the proposed MINMPC controller.

[0019] In some embodiments of the invention, the number of MITCP iterations is determined by the termination condition of the optimization algorithm, which may be based on the norm of the Karush-Kuhn-Tucker (KKT) necessary condition of optimality for the MINLP, excluding the integer condition.

[0020] In some embodiments of the invention, a homotopy-based penalty method may be used to adjust the cost function in each MICP subproblem to progressively implement the MISCP algorithm to compute updates to optimal solution guesses for some or all of the integer and / or binary decision variables that remain close to the solution guesses for the integer and / or binary decision variables in the previous MISCP iteration.

number

[0021] Some embodiments of the present invention are based on the recognition that the use of homotopy-type penalty methods for some or all of the integer and / or binary decision variables can prevent cycles in the MISCP optimization algorithm. In addition, some embodiments of the present invention are based on the recognition that the use of homotopy-type penalty methods for some or all of the integer and / or binary decision variables can significantly reduce the computational cost of solving the MICP subproblem in the MISCP optimization algorithm.

[0022] Accordingly, one embodiment discloses a predictive feedback controller for controlling a hybrid dynamic system with nonlinear dynamics and continuous and discrete elements of motion, the predictive feedback controller comprising at least one processor and a memory storing instructions that, when executed by the at least one processor, cause the predictive feedback controller to: ● accepting a feedback signal including a measurement indicative of a current state of the hybrid dynamic system, including one or a combination of a current state of the predictive controller, a current state of one or more actuators of the hybrid dynamic system, and a current state of an output of the hybrid dynamic system; ● formulating a mixed integer nonlinear programming (MINLP) problem that optimizes an objective function subject to one or more constraints, with a solution that indicates control commands for modifying a current state of the hybrid dynamic system according to a control objective, the constraints including equality constraints, inequality constraints, or both, the constraints and the control objective of the MINLP problem including one or more nonlinear functions of continuous optimization variables that represent continuous elements of the hybrid dynamic system's behavior and one or more linear functions of integer optimization variables that represent discrete elements of the hybrid dynamic system's behavior, the MINLP problem being formulated into a separable format that ensures that the discrete elements of the behavior exist only in linear functions of the MINLP problem; further, the instructions, when executed by the at least one processor, cause the predictive feedback controller to: ● solving the MINLP problem over multiple iterations of a sequential convexification-based optimization procedure until a termination condition is met, to perform the iterations, the predictive feedback controller is configured to perform a partial convexification of a portion of a space of solutions that includes a current solution guess, the partial convexification generating a convex approximation of a nonlinear function of the MINLP without approximating a linear function of the MINLP to generate the partially convexified MINLP; further, the instructions, when executed by the at least one processor, cause the predictive feedback controller to update the current solution guess by solving a mixed-integer convex programming (MICP) formulation of the partially convexified MINLP problem; • Submitting control commands generated according to the solution of the MINLP problem to the hybrid dynamic system, thereby causing a change in the current state of the hybrid dynamic system. [Brief description of the drawings]

[0023] [Figure 1A]FIG. 1 illustrates a block diagram of a mixed integer nonlinear model predictive controller and feedback system according to some embodiments. [Figure 1B] FIG. 1 illustrates a block diagram of a mixed integer nonlinear model predictive controller and feedback system according to some embodiments. [Figure 1C] FIG. 1 illustrates a block diagram of a hierarchical integration between a mixed integer nonlinear model predictive controller and a tracking controller for controlling a feedback system, according to some embodiments. [Figure 2A] FIG. 1 illustrates a block diagram of a method for mixed integer nonlinear model predictive control (MINMPC) to implement a predictive controller that calculates control signals given current states and control commands of a system, according to some embodiments. [Figure 2B] FIG. 1 illustrates a block diagram of a MINMPC method for solving optimally controlled structured mixed integer nonlinear programs (MINLPs) according to some embodiments. [Figure 2C] 1 illustrates a reformulation into a separable format of the MINLP solved in an implementation of the MINMPC method, according to some embodiments. [Figure 2D] FIG. 1 shows a block diagram of a MINMPC method for solving optimal control structured MINLPs in a separable format according to some embodiments. [Figure 3A] FIG. 1 shows a block diagram of an iterative optimization procedure according to some embodiments, where each iteration consists of a partial convexification step and the solution of a mixed-integer convex programming (MICP) subproblem to compute a feasible and (locally) optimal solution for the MINMPC. [Figure 3B] FIG. 1 shows a block diagram of an iterative optimization procedure according to some embodiments, where each iteration consists of a partial convexification step and the solution of a MICP subproblem subject to trust-region constraints to compute a feasible and (locally) optimal solution for MINMPC. [Figure 4A] FIG. 1 illustrates a block diagram of one or more iterations of a sequential convexification-based optimization procedure for computing a feasible and (locally) optimal MINLP solution according to some embodiments of the present invention. [Figure 4B] FIG. 1 illustrates a block diagram of one or more iterations of a sequential convexification-based optimization procedure according to some embodiments, where each iteration consists of a local linear-quadratic approximation in a small region of the MINLP solution space and the solution of an MIQP subproblem. [Figure 4C] FIG. 1 shows a block diagram of a sequential convexification-based optimization procedure for computing a feasible and (locally) optimal MINLP solution according to some embodiments of the present invention, where the values ​​of integer variables are fixed after one or more iterations. [Figure 5A] FIG. 1 shows a schematic diagram of a partial convexification step in which one or more convex inequality constraints are computed to approximate one or more non-convex inequality constraints in a small region of the MINLP solution space in a local neighborhood of the current solution guess, according to some embodiments. [Figure 5B] FIG. 1 shows a schematic diagram of a piecewise convexification step that computes a convex approximation of one or more smooth nonlinear functions in a cost function of a MINLP in a separable format, according to some embodiments. [Figure 6A] FIG. 1 illustrates a block diagram of a compact formulation of an optimal control structured MINLP in a separable format that is solved at each time step of a MINMPC controller according to some embodiments of the present invention. [Figure 6B] FIG. 1 shows a block diagram of a sequential convexity-based optimization procedure for updating a MINLP solution guess based on local linear-quadratic approximations in small regions of the MINLP solution space and solutions to MIQP subproblems, according to some embodiments. [Figure 6C] FIG. 1 shows a block diagram of a sequential convexification-based optimization procedure that fixes the values ​​of integer variables after one or more iterations and solves the convex QP subproblem in one or more subsequent iterations to update the MINLP solution guess. [Figure 6D] FIG. 1 shows a block diagram of a sequential convexification-based optimization procedure for computing a solution to a MINMPC problem, in which each iteration consists of a partial convexification step followed by the solution of a convex QP or non-convex MIQP subproblem, according to some embodiments. [Figure 7A]FIG. 1 illustrates a block diagram of a merit function for quantifying optimality and constraint satisfaction for a MINLP solution guess in a sequential convexification-based optimization procedure, according to some embodiments. [Figure 7B] FIG. 1 shows a block diagram of a line search procedure for calculating a step size based on one or more merit function evaluations and directional derivatives for updating a MINLP solution guess in a sequential convexification-based optimization procedure, according to some embodiments. [Figure 7C] 1 shows pseudocode for a MISQP optimization method based on a line search procedure for computing a feasible and (locally) optimal solution for a MINLP in a MINMPC controller, according to some embodiments. [Figure 8A] FIG. 1 shows a block diagram of the ratio of actual versus predicted reduction in merit function values ​​at each iteration of a sequential convexity-based optimization procedure, in accordance with some embodiments. [Figure 8B] 1 shows a block diagram of a trust region procedure that updates a MINLP solution guess at each iteration of a sequential convexification-based optimization procedure based on the ratio of actual versus predicted reduction in the value of a merit function and based on a trust region radius value. [Figure 8C] FIG. 1 shows a block diagram of a MISQP optimization algorithm based on a partial convexity step to construct and solve MIQP subproblems subject to trust region constraints to limit the search direction within a small region of the MINLP solution space, according to some embodiments. [Figure 8D] 1 shows pseudocode for a MISQP optimization method, based on a trust region search procedure, for computing a feasible and (locally) optimal solution for a MINLP in a MINMPC controller, according to some embodiments. [Figure 9A] We show an example of a homotopy-type penalty method by adding one or more additional penalty terms to the MINLP objective function, which can be used to accelerate the convergence of a sequential convexification-based optimization procedure according to some embodiments of the present invention. [Figure 9B]We show a schematic diagram of an example of a homotopy-type penalty method, which can be used to accelerate the convergence of the MISCP optimization algorithm to find a feasible and (locally) optimal solution for the MINLP by adding one or more additional penalty terms to the MINLP objective function. [Figure 9C] We show a schematic diagram of an example of a homotopy-type penalty method by adding one or more additional linear and / or smooth nonlinear inequality constraints, which can be used to accelerate the convergence of the MISCP optimization algorithm to find a feasible and (locally) optimal solution for the MINLP. [Figure 10A] FIG. 2 shows a block diagram of a warm-start initialization procedure for computing an initial MINLP solution guess given an (approximate) solution to a MINMPC problem at a previous control time step, according to some embodiments of the present invention. [Figure 10B] FIG. 1 shows a block diagram of a warm-start initialization procedure, based on a time-shifting operation, for computing an initial solution guess for an optimal control-structured MINLP given an (approximate) solution to a MINMPC problem at the previous control time step, according to some embodiments. [Figure 11A] FIG. 2 illustrates a schematic diagram of an example binary control-variable search tree that represents a nested tree of search spaces for integer-feasible solutions to MICP subproblems, according to some embodiments. [Figure 11B] FIG. 2 illustrates a block diagram of a branch-and-bound mixed-integer optimization algorithm for searching integer-feasible optimal solutions to MICP subproblems based on a nested tree of search spaces and corresponding lower / upper bounds, according to some embodiments. [Figure 12A] 1 shows a schematic diagram of a vehicle including a predictive controller employing principles of some embodiments. [Figure 12B] FIG. 2 illustrates a schematic diagram of the interaction between a predictive controller and other controllers of the vehicle, according to some embodiments. [Figure 12C] 1 shows a schematic diagram of a path and / or motion planning method for a controlled vehicle employing principles of some embodiments. [Figure 12D]1 illustrates an example traffic scene for a single-vehicle or multi-vehicle decision-making module according to some embodiments. [Figure 13A] 1 shows a schematic diagram of a spacecraft mixed-integer predictive control problem formulation employing principles, according to some embodiments. [Figure 13B] 1 shows a schematic diagram of a spacecraft mixed-integer predictive control problem formulation employing principles, according to some embodiments. [Figure 14A] 1 shows a schematic diagram of a vapor compression system controlled by a controller according to some embodiments. [Figure 14B] 1 illustrates an example of an arrangement of signals, sensors, and controllers used within a VCS, according to some embodiments. [Figure 15] 1 illustrates a method for controlling a system according to an example embodiment. DETAILED DESCRIPTION OF THE PREFERRED EMBODIMENTS

[0024] Some embodiments of the present disclosure provide systems and methods for controlling the operation of a system or systems that use a predictive controller. An example of a predictive controller is a mixed integer nonlinear model predictive controller (MINMPC), which determines control inputs based on a nonlinear model of a controlled system having continuous and discrete elements of operation.

[0025] FIG. 1A illustrates a block diagram of a predictive controller 110 and a feedback system 120 according to some embodiments. FIG. 1A illustrates an example feedback system (or system) 120 connected to the predictive controller 110 (or controller) via a state estimator 130 according to some embodiments. In some implementations, the predictive controller 110 is a MINMPC controller programmed according to a dynamic model 102 (or system model) of the system 120. The system model 102 can be a set of equations that represent the changes in state and output 103 of the system 120 as a function of current and previous inputs 111 and previous outputs 103 over time. The system model 102 can include constraints 104 that represent physical and operational limitations of the system 120. During operation, the controller 110 receives commands 101 that indicate a desired behavior of the system 120. The commands can be, for example, motion commands. In response to receiving the commands 101, the controller 110 generates control signals 111 that serve as inputs to the system 120, which includes both continuous and discrete elements of operation. In response to the input, the system updates the output 103 of the system 120. Based on the measurements of the output 103 of the system 120, the estimator 130 updates an estimated state 121 of the system 120. This estimated state 121 of the system 120 provides state feedback to the predictive controller 110. Thus, the predictive controller first receives a feedback signal 121 of the system 120 via the estimator 130, where the feedback signal 121 includes measurements of the state of the system 120.

[0026] The system 120 referred to herein may be any machine or device controlled by a particular operational input signal, e.g., a control signal 111 (input). The control input signal may possibly include continuous elements, such as voltage, pressure, force, torque, steering angle, speed, and temperature, as well as discrete elements, such as energy levels, quantized valve inputs, gear shifting, on / off actuation, lane selection, and obstacle avoidance decision variables. The system 120 returns a number of controlled output signals 103 (output), possibly including continuous elements, such as current, flow, speed, position, temperature, heading, and steering angle, as well as discrete elements, such as energy levels, quantized valve states, gear states, on / off states, and lane positions. The output values ​​are partly related to previous output values ​​of the system and partly related to previous and current input values. The dependency on previous inputs and previous outputs is encoded in the state of the system. The operation of the system, e.g., the movement of the components of the system, may include a sequence of output values ​​generated by the system after application of certain input values.

[0027] The system model 102 may include a set of mathematical equations that describe how the system output changes over time as a function of current and previous inputs and previous outputs. The mathematical equations may include one or more smooth equations that depend on continuous variables and one or more mixed integer equations that depend on both continuous and discrete variables. Each function in the mathematical equations may be either linear or a smooth nonlinear function. The state of the system 120 is any set of information that is generally time-varying, such as, for example, a proper subset of current inputs and outputs and previous inputs and outputs that, together with the model and future inputs of the system, can uniquely define the future motion of the system.

[0028] The system 120 may be subject to physical limitations and specification constraints 104 that limit the range over which the outputs, inputs, and possibly also states of the system 120 may be operated. The constraints may include one or more smooth equations that depend on continuous variables and one or more mixed integer equations that depend on both continuous and discrete variables. The constraint functions may be either linear or smooth nonlinear functions. Some embodiments of the present invention are based on the recognition that first or higher order directional derivatives may be calculated for each of the smooth nonlinear functions in the constraints 104.

[0029] The predictive controller 110, which may be implemented in hardware or as a software program running on a processor, e.g., a microprocessor, receives the estimated state 121 and the desired motion commands 101 of the system 120 at fixed or variable control period sampling intervals, and uses this information to determine inputs, e.g., control signals 111, for operating the system 120. The predictive controller 110 further uses a mixed-integer sequential convex programming (MISCP) method to solve an optimal control-structured mixed-integer nonlinear programming (MINLP) problem and optimizes a current solution of an objective function subject to one or more equality and / or inequality constraints over multiple iterations until a termination condition is met, e.g., the solution is feasible and (locally) optimal for the optimal control-structured MINLP. Each iteration of the MISCP method performs a partial convexification of a portion of the space of solutions that includes the current solution, and the partial convexification produces a convex approximation of a smooth nonlinear function of the MINLP without approximating the linear function of the MINLP to produce the partially convexified MINLP. Each iteration may then update the current solution by solving a mixed integer convex programming (MICP) formulation of the partially convexified MINLP problem to generate control signals 111. Furthermore, the predictive controller 110 controls the system 120 based on the control signals 111 to change the state of the system 120.

[0030] The estimator 130, which can be implemented in hardware or as a software program running on either the same or a different processor as the controller 110, receives the output of the system 103 at fixed or variable control period sampling intervals and uses new and previous output measurements to determine an estimated state 121 of the system 120.

[0031] Thus, by using the MISCP optimization method, the processor calculates a feasible and (locally) optimal solution of the optimal control structured MINLP by solving a sequence of MICP subproblems, where the formulation of each MICP subproblem is based on a partial convexification of a portion of the solution space. The reduced computational complexity of solving the MICP subproblems compared to the computational complexity of solving the original MINLP enables the processor to accurately determine the feasible and (locally) optimal solution for controlling the state of the system 120 in less time. Thus, the processor achieves high processing speed with high accuracy.

[0032] 1B shows a block diagram of a mixed integer nonlinear predictive controller 110 and a feedback system 120 according to some embodiments. The predictive controller 110 operates the system 120 such that the estimated states 121 and outputs 103 of the system 120 comply with the commands 101. The controller 110 includes a computer, for example in the form of a single central processing unit (CPU) or multiple CPU processors 151 connected to a memory 152 for storing the system model 102 and constraints 104 on the operation of the system 120. The CPU processor 151 may be comprised of a single core processor, a multi-core processor, a computing cluster, or any number of other configurations. The memory 152 may include a random access memory (RAM), a read only memory (ROM), a flash memory, or any other suitable memory system.

[0033] 1C shows a block diagram of a hierarchical integration between a mixed integer nonlinear predictive controller 110 that calculates a high-level control signal 111 and a tracking controller 115 that aims to track the high-level control signal 111 and calculate a low-level control signal 112 to directly control a feedback system 120, according to some embodiments of the present invention. For example, the predictive controller 110 can be a MINMPC controller that calculates a reference motion trajectory based on a nonlinear dynamic model of the controlled system having continuous and discrete elements of motion, and the tracking controller 115 aims to execute (a part of) the reference motion trajectory by sending control inputs 112 directly to the actuators of the controlled system.

[0034] An example of the tracking controller 115 may be based on a proportional-integral-derivative (PID) controller for tracking the time-varying reference motion trajectory calculated by the MINMPC controller 110. In some embodiments of the present invention, the tracking controller 115 may be based on a model predictive controller (MPC) with a relatively simplified dynamic model and a simplified set of constraints, and therefore requires relatively low computational complexity. For example, the MPC tracking controller 115 may be based on a linear-quadratic objective function, a linear dynamic model, and linear equality and inequality constraints, which require the solution of a convex linear programming (LP) or convex quadratic programming (QP) problem that is computationally much easier to solve than the MINLP solved by the MINMPC controller 110.

[0035] Some embodiments of the present invention are based on the recognition that a relatively long prediction horizon may be used for the mixed integer nonlinear predictive controller 110 and a relatively short prediction horizon may be used for the tracking controller 115, and a relatively fast sampling rate may be used for the tracking controller, e.g., with a sampling period of 10-100 milliseconds, while a relatively slow sampling rate may be used for the mixed integer nonlinear predictive controller 110, e.g., with a sampling period of 0.5-2 seconds.

[0036] Some embodiments of the present invention are based on the recognition that due to the computational complexity of solving a MINLP, a relatively high-order, low-precision dynamic model 102 and constraints 104 can be used in the MINLP formulation of the mixed integer nonlinear predictive controller 110, while a relatively low-order, high-precision dynamic model 102 and constraints 104 can be used in the design and formulation of a computationally cheaper tracking controller 115. [What can be the minimum ratio of long-term / short-term prediction horizons? How can partial convexity help to achieve this ratio?]

[0037] 2A illustrates a block diagram of a system and method for mixed integer nonlinear model predictive control (MINMPC) to implement a predictive controller 110 that calculates a control signal 111 given a current state and control commands 101 of the system 121, according to some embodiments. Specifically, MINMPC calculates a control solution, e.g., a solution vector 255 that may include a sequence of future optimal discrete and continuous control inputs over a forecast time horizon of the system 120, by solving a constrained mixed integer nonlinear programming (MINLP) problem 250 at each control time step. The MINLP data 245 of the objective function, equality constraints, and discrete, continuous, and mixed integer inequality constraints in this optimization problem 250 depend on the dynamic model, system constraints 240, the current state of the system 121, the control objective, and the control commands 101.

[0038] In some embodiments of the invention, the solution of this inequality constrained MINLP problem 250 uses one or more state and control values ​​over the forecast time horizon, and potentially other MINLP solution information from the previous control time step, which can be read from memory (210). This concept is called a warm or hot start of the optimization algorithm, and in some embodiments, can reduce the required computational effort of the MINMPC controller. Similarly, the corresponding solution vector 255 can be used to update and store in memory (235) one or more optimal sequences of state and control values ​​over the forecast time horizon, and potentially other MINLP solution information for the next control time step.

[0039] In some embodiments, the mixed integer optimization algorithm is based on a search algorithm for solving the MICP subproblem that is the result of the partial convexification step in each iteration of the sequential convexification algorithm, and the MINMPC controller updates and stores additional mixed integer programming solution information 260 to reduce the computational effort of the search algorithm in one or more iterations in the current and / or next control time step. In one embodiment, the MICP subproblem in each iteration is solved using a branch and bound optimization method, and the warm start information may include data related to nodes in a binary search tree that are part of the solution path from the root node to the leaf node where the optimal integer feasible control solution is found, in order to improve the node selection and variable branching strategy from one iteration to the next.

[0040] 3A shows a block diagram of the MINMPC method for solving an optimal control structured MINLP 250 to calculate the control signal 111 at each control time step given the current state 121 and control commands 101 of the system 121. Some embodiments of the invention are based on a nonlinear dynamic model of the system 263 with linear equality constraints 262, nonlinear mixed integer inequality constraints 264 at each time step within the prediction time horizon, nonlinear mixed integer inequality constraints 265 at the terminal time steps of the prediction time horizon, linear discrete equality constraints 266, and a linear-quadratic or nonlinear objective function 261, and the resulting constrained MINLP 260 needs to be solved at each control time step. The MINLP data 245 can include MINLP matrices and vectors 246 and MINLP functions 247 to formulate the optimal control structured MINLP 260.

[0041] In some embodiments of the invention, the nonlinear dynamic model 263 of the system is described by one or more linear and / or smooth nonlinear differential equations. In some embodiments of the invention, the dynamic model 263 of the system describes a linear or nonlinear hybrid system with state-dependent and / or input-dependent jumps in the dynamic equations, including, for example, piecewise linear and / or piecewise smooth nonlinear equations.

number

[0042]

number

[0043]

number

[0044] 2C illustrates the reformulation of the optimal control structured MINLP 260 into a separable format, i.e., with separation of smooth nonlinear functions of continuous and integer variables that enter the MINLP linearly, possibly by adding one or more continuous and / or integer variables (270). In some embodiments, the MINMPC predictive controller 110 solves (280) the resulting MINLP in a separable format to compute the control signals 111 at each control time step given the current state 121 and commands 101 of the system 120.

[0045] In some embodiments of the invention, the optimal control structured MINLP 260 can be trivially reformulated 270 in a separable format because, for example, we can define the function 261 in the objective, the function 263 in the equality constraints, and the functions 264-265 in the inequality constraints as follows:

number

[0046]

number

[0047] Some embodiments of the invention are based on the recognition that MINLP 280 in separable format is mathematically equivalent to the original MINLP 260, i.e., MINLP 280 in separable format is infeasible if and only if the original MINLP 260 is infeasible, and that a feasible (locally) optimal solution to MINLP 280 in separable format can be used to construct a feasible (locally) optimal solution to the original MINLP 260. For example, in some embodiments of the invention, an affine or nonlinear transformation exists to convert a feasible (locally) optimal solution to MINLP 280 in separable format to a feasible (locally) optimal solution of the original MINLP 260.

[0048] 2D shows a block diagram of the MINMPC method for solving an optimal control structured MINLP 280 in a separable format to calculate the control signal 111 at each control time step given the current state 121 and commands 101 of the system 120. Some embodiments of the invention are based on a nonlinear dynamic model 283 of the system with linear equality constraints 282, nonlinear mixed integer inequality constraints 284 at each time step within the prediction time horizon, nonlinear mixed integer inequality constraints 285 at the terminal time steps of the prediction time horizon, linear discrete equality constraints 286, and a linear-quadratic or nonlinear objective function 281, and the resulting constrained MINLP 280 needs to be solved at each control time step. The MINLP data 245 can include MINLP matrices and vectors 246 and MINLP functions 247 to formulate the optimal control structured MINLP 280.

[0049] For example, MINLP matrices and vectors 246 are expressed as matrices D k , E k , E N and vector c k may include.

number

[0050] In some embodiments of the invention, the nonlinear dynamic model 283 of the system is described by one or more linear and / or smooth nonlinear differential equations. In some embodiments of the invention, the dynamic model 283 of the system describes a linear or nonlinear hybrid system with state-dependent and / or input-dependent jumps in the dynamic equations, including, for example, piecewise linear and / or piecewise smooth nonlinear equations.

number

[0051]

number

[0052]

number

[0053] FIG. 3A shows a block diagram of an iterative optimization procedure for solving an optimal control structured MINLP 250, where each iteration consists of a partial convexification step 315 and solution 320 of a mixed integer convex programming (MICP) subproblem, updating intermediate solution guesses for integer variables 325 and continuous variables 330 until a feasible and (locally) optimal solution is found (255) that can be used to construct control signals 111 at each control time step, given the current state 121 and commands 101 of the system 120.

number

[0054] Each iteration of the sequential convexification-based optimization procedure involves checking 350 whether the current intermediate solution guess is feasible and / or (locally) optimal, in which case either a solution is found (255) or a new iteration is performed to update the current intermediate MINLP solution guess 335 based on the partial convexification step 315 and the MICP subproblem solution 320, and the current iteration number k may be updated (340). In some embodiments of the invention, the termination condition 350 involves checking whether both current intermediate solution guesses are feasible, i.e., whether the solution satisfies all equality and inequality constraints in the MINLP, whether it is close enough to a globally optimal solution, and / or whether the computational cost of the iterative optimization procedure has reached a certain time limit. For example, a maximum number of iterations can be imposed to ensure that the iterative optimization procedure terminates with a deterministic runtime.

[0055] Some embodiments of the present invention are based on the recognition that each iteration of the sequential convexification-based optimization procedure that solves 270 the separable-format MINLP can be based on including a partial convexification step 315 of only the smooth nonlinear part of the separable-format MINMPC problem, i.e., for each of the nonlinear equality and inequality constraints and / or for each of the nonlinear objective functions. The partial convexification step 315 is based on linear functions that depend on one or more integer variables, e.g., the matrix D in the constraint and objective functions of the optimal control structured MINLP 280. k , E k , E N and vector c k In particular, the partially convexified step 315 does not change the linear constraint and objective functions in the separable MINLP format 280 based on k ,Δz k ) (320).

[0056] Some embodiments of the present invention are based on the recognition that, depending on the particular implementation of the partial convexification step 315, a MICP subproblem may correspond, for example, to a mixed integer linear programming (MILP) subproblem, a mixed integer quadratic programming (MIQP) subproblem, a mixed integer quadratically constrained quadratic programming (MIQCQP) subproblem or a mixed integer quadratic cone programming (MISOCP) subproblem.

[0057]

number

[0058] FIG. 3B shows a block diagram of an iterative optimization procedure for solving the optimal control structured MINLP 250, where each iteration consists of a partial convexification step 315 and solution 360 of a mixed-integer convex programming (MICP) subproblem subject to trust region constraints, updating intermediate solution guesses for integer and continuous variables (335) until a feasible and (locally) optimal solution is found (255) that can be used to construct control signals 111 at each control time step, given the current state 121 and commands 101 of the system 120.

number

[0059]

number

[0060]

number

[0061]

number

[0062]

number

[0063] Figure 4B shows the feasible and (locally) optimal MINLP solution (y N* , z N* 4 shows a block diagram of one or more iterations of a successive convexity-based optimization procedure for computing a MINLP function (MPC function) 255, where the partial convexity steps 401-402 in each iteration are based on local linear-quadratic approximations 441-442 in a small region of the MINLP solution space, and the MICP subproblems 411-412 in each iteration correspond to solutions 451-452 of the MIQP subproblems for computing search directions in each iteration of the optimization procedure. In some embodiments of the invention, the partial convexity steps in each iteration are based on local linearization of only the smooth nonlinear parts of the MINMPC problem in a separable format, i.e., including a local linearization step for each of the nonlinear equality and inequality constraints and / or a local linear-quadratic approximation step for each of the nonlinear objective functions.

[0064]

number

[0065] Some embodiments of the invention are based on the recognition that solutions 453 of convex QP subproblems are generally computationally much cheaper than solutions 451-452 of non-convex MIQP subproblems at each iteration of the sequential convexification-based optimization procedure. In some embodiments of the invention, the decision to fix the values ​​of all integer variables 460 is based on whether a current intermediate solution guess is close enough to a globally optimal solution and / or whether the computational cost of the iterative optimization procedure has reached a certain time limit. For example, in some embodiments, a maximum number N miqp For example, k ≥ Nmiqp After iteration 460, the integer variable Δz k By fixing the value of =0, the computational effort of the iterative optimization procedure can be significantly reduced.

[0066] 5A shows a schematic diagram of a partial convexification step that computes one or more convex inequality constraints to approximate one or more non-convex inequality constraints in a small region of the MINLP solution space corresponding to a local neighborhood of a current intermediate solution guess. Some embodiments of the invention are based on the recognition that the partial convexification step merely computes a convex approximation for one or more smooth nonlinear inequality constraint functions in MINLP 280 in a separable format, while keeping invariant linear functions that may depend on one or more integer variables.

[0067]

number

[0068] In some embodiments of the present invention, each iteration of the sequential convexification-based optimization procedure performs a partial convexification step to compute a convex approximation 525 of one or more non-convex constraints 510 in a local neighborhood of the current solution guess 520 in order to compute a search direction that is approximately inside 526 of the non-convex feasible region 511. Some embodiments of the present invention may be further modified so that each iteration of the sequential convexification-based optimization procedure performs a partial convexification step to compute a convex approximation 525 of one or more non-convex constraints 510 in a local neighborhood of the current solution guess 520 in order to compute a search direction that is approximately inside 526 of the non-convex feasible region 511. k ∈{0,1} (505), thereby computing a solution to a MICP subproblem 320 that is non-convex due to one or more integer variables, including k = 0 (506), while a different convex approximation 535 of one or more non-convex constraints 515 exists for the value z k = 1 (507) may be in the local neighborhood of the current solution guess 530.

[0069] For example, the solution to the MICP subproblem 320 can be calculated by multiplying the binary variables z by the corresponding convex approximation 525 or 535 of the smooth nonlinear part of the MINLP. k ∈{0,1}(505) to z k =0(506) or z k Some embodiments of the present invention may include setting the binary decision variable z k Depending on the value of ∈{0,1}(505), i.e., z k =0(506) or z k = 1 (507), for each iteration of the sequential convexification-based optimization procedure, there may be a complex transformation 540 between the current solution guess 520 or 530, the non-convex feasible region 511 or 516, and the corresponding convex approximation 525 or 535.

[0070] In some embodiments of the invention, the partial convexification step in each iteration of the sequential convexification-based optimization procedure is based on a local linearization of one or more smooth nonlinear inequality constraint functions that define a non-convex set of feasible values ​​511 or 516. Specifically, the partial convexification step may compute one or more linear inequality constraints, i.e., to remain within one or more half-spaces 526 with respect to one or more linear functions 525, to approximate the non-convex set of feasible values ​​511 defined by one or more smooth nonlinear inequality constraints 284-285 within a subregion of the MINLP solution space corresponding to a local neighborhood of the current intermediate solution guess 520. Similarly, the partial convexification step may compute one or more linear equality constraints to approximate the non-convex set of feasible values ​​defined by one or more smooth nonlinear equality constraints 283 within a subregion of the MINLP solution space corresponding to a local neighborhood of the current intermediate solution guess.

[0071]

number

[0072] In some embodiments of the invention, each iteration of the sequential convexification-based optimization procedure performs a partial convexification step that computes a convex approximation 565 of one or more non-convex cost functions 550 in a local neighborhood of a current solution guess 560 in order to compute a search direction that is approximately towards a local minimum 561 or a global minimum 562 of the non-convex cost function 550. Some embodiments of the invention may be further modified so that each iteration of the sequential convexification-based optimization procedure performs a partial convexification step that computes a convex approximation 565 of one or more non-convex cost functions 550 in a local neighborhood of a current solution guess 560 in order to compute a search direction that is approximately towards a local minimum 561 or a global minimum 562 of the non-convex cost function 550. k ∈{0,1} 505, thereby computing a solution to a MICP subproblem 320 that is non-convex due to one or more integer variables, k = 0 (506), while a different convex approximation 575 of one or more non-convex cost functions 555 exists for the value z k = 1 (507) may be in the local neighborhood of the current solution guess 570.

[0073] Some embodiments of the present invention are directed to a binary decision variable z k Depending on the value of ∈{0,1}(505), i.e., z k =0(506) or z k = 1 (507), for each iteration of the sequential convexity-based optimization procedure, there may be a complex transformation 580 between the current solution guess 560 or 570, the non-convex cost function 550 or 555, and the corresponding convex approximation 565 or 575.

[0074] In some embodiments of the invention, the partial convexification step in each iteration of the sequential convexification-based optimization procedure is based on a local linear or linear-quadratic approximation 565 or 575 of one or more smooth nonlinear objective functions 281 that define the nonconvex cost function 550 or 555. Specifically, the partial convexification step may calculate one or more linear or linear-quadratic objective functions 565 or 575 to approximate the nonconvex cost function 545 within a small region of the MINLP solution space corresponding to a local neighborhood of the current intermediate solution guess 560 or 570, and the resulting MICP subproblem includes the one or more linear or linear-quadratic objective functions 565 or 575 such that a solution of the MICP subproblem defines a search direction in the MINLP solution space toward a local or global minimum of the nonconvex cost function 545.

[0075] 6A shows a block diagram of a compact formulation of an optimal control structured MINLP in a separable format that is solved at each time step of a MINMPC controller according to some embodiments of the present invention. Specifically, the original optimal control structured MINLP calculates the matrix D k , E k , E N and vector c k Based on this, in a separable format, the function includes a function in the objective 281, a function in the equality constraints 283, and functions in the inequality constraints 284-285.

number

[0076]

number

[0077]

number

[0078] FIG. 6B shows a block diagram of an implementation of a sequential convexification-based optimization procedure that updates a MINLP solution guess based on local linear-quadratic approximations and solutions of MIQP subproblems in small regions of the MINLP solution space, according to some embodiments of the present invention.

number

[0079] In some embodiments of the present invention, the MIQP subproblem 650 is constructed based on a sequential convex programming (SCP) or sequential quadratic programming (SQP) method applied to smooth nonlinear functions in the compact formulation 630 of the MINLP in a separable format with respect to the continuous variable y 636, i.e., without affecting the linear terms with respect to the integer variable z 637 of the MINLP.

number

[0080]

number

[0081]

number

[0082]

number

[0083]

number

[0084]

number

[0085]

number

[0086]

number

[0087] FIG. 6D shows a block diagram of an iterative sequential convexification-based optimization procedure for solving the optimal control structured MINLP 250, similar to the block diagram of FIG. 3A, where each iteration consists of a partial convexification step 315 followed by the solution of a convex QP subproblem 675 or MIQP subproblem 680 that updates intermediate solution guesses for integer variables 325 and continuous variables 330 until a feasible and (locally) optimal solution is found (255), which can be used to construct the control signal 111 at each control time step given the current state 121 and commands 101 of the system 120.

number

[0088]

number

[0089]

number

[0090]

number

[0091]

number

[0092]

number

[0093] In some embodiments of the invention, each iteration of the MISCP optimization algorithm calculates a parameter value ρ>0 that is large enough so that the descent condition 721 holds. Embodiments of the invention are based on the recognition that increasing the parameter value ρ>0 results in a merit function 711 that quantifies the constraint satisfaction relatively more strongly compared to the optimality of the MINLP solution guess. Alternatively, decreasing the parameter value ρ>0 results in a merit function 711 that quantifies the optimality relatively more strongly compared to the constraint satisfaction for the MINLP solution guess.

[0094] FIG. 7B shows a block diagram of a line search procedure for computing a step size based on one or more merit function evaluations and directional derivative calculations to update a MINLP solution guess (730) in the sequential convexification-based optimization procedure described in FIG. 3A.

number

[0095]

number

[0096] In some embodiments of the present invention, an iterative backtracking procedure is used to determine the initial value α k Starting from =1, step size value 0≦α k 1, and this step size value is increased by 0≦α k can be iteratively reduced towards

[0097]

number

[0098]

number

[0099]

number

[0100]

number

[0101] FIG. 8A shows the MINLP solution guess (y k ,z k ) and the MICP subproblem solution (Δy k ,Δz k ), the ratio of actual reduction to predicted reduction in the value of the merit function 810, R k Given the compact formulation of the objective and constraint functions in MINLP 705, a merit function can be defined to quantify the optimality and constraint satisfaction of the MINLP solution guess at each MISCP iteration (711),

number

[0102]

number

[0103]

number

[0104]

number

[0105]

number

[0106]

number

[0107] FIG. 8C shows a block diagram of a trust region procedure for determining how to update the MINLP solution guess in the sequential convexification-based optimization procedure described in FIG. 3B, which builds and solves (360) a MICP subproblem subject to trust region constraints to limit the search direction within a small region of the MINLP solution space at each iteration of the MISCP optimization algorithm, based on the partial convexification step 315.

number

[0108] Some embodiments of the present invention generally consider a MICP subproblem with a trust region radius d k This is based on the realization that it becomes computationally cheaper to solve for smaller values ​​of . Therefore, in some embodiments of the invention, the trust region procedure aims to reduce the trust region radius when possible without slowing down the convergence of the MISCP optimization algorithm to a feasible and (locally) optimal MINLP solution.

[0109]

number

[0110]

number

[0111]

number

[0112]

number

[0113] 9A illustrates an example of a homotopy-based penalty method that can be used to accelerate the convergence of a sequential convexification-based optimization procedure in accordance with some embodiments of the present invention. For example, in some embodiments of the present invention, an integer variable in MINLP 630 is a function of one or more binary optimization variables z j ∈{0,1} 900, one or more additional penalty terms may be added to the MINLP objective function 631 to avoid cycling between different values ​​of the MINLP solution guess in subsequent iterations of the MICP optimization algorithm. Some embodiments of the present invention are based on the recognition that adding one or more additional penalty terms to the MINLP objective function 631 can result in a MICP subproblem 650 that is computationally cheaper to solve in each iteration of the MICP optimization algorithm.

[0114]

number

[0115] 9B shows a schematic diagram of an example of a homotopy-based penalty method that can be used to accelerate the convergence of the MISCP optimization algorithm to find a feasible and (locally) optimal solution of the MINLP. In some embodiments of the invention, the homotopy-based penalty method can be used to adjust the cost function in each MICP subproblem 320 to progressively implement the MISCP optimization algorithm to compute updates to the MINLP solution guesses for some or all of the integer and / or binary decision variables that remain close to the solution guesses for the integer and / or binary decision variables in the previous MISCP iteration.

number

[0116]

number

[0117] Some embodiments of the present invention are based on the recognition that the use of homotopy-type penalty methods for some or all of the integer and / or binary decision variables can prevent cycles in the MISCP optimization algorithm. In addition, some embodiments of the present invention are based on the recognition that the use of homotopy-type penalty methods for some or all of the integer and / or binary decision variables can significantly reduce the computational cost of solving the MICP subproblem in the MISCP optimization algorithm. For example, some embodiments of the present invention are based on a branch-and-bound method to solve the MICP subproblem in each iteration of the MISCP optimization algorithm, and adding one or more linear penalty terms to the cost function can result in a significantly smaller branch-and-bound search tree, and therefore the computational cost of solving the MICP subproblem can be significantly reduced.

[0118] 9C shows a schematic diagram of an example of a homotopy-based penalty method by adding one or more additional linear and / or smooth nonlinear inequality constraints that can be used to accelerate the convergence of the MISCP optimization algorithm to find a feasible and (locally) optimal solution of the MINLP. For example, in some embodiments of the present invention, the integer variables of the MINLP 630 are expressed as one or more binary optimization variables z j∈{0,1} 900, one or more additional linear and / or smooth nonlinear inequality constraints may be added to the MINLP constraints 633 to avoid cycling between different values ​​of the MINLP solution guess in subsequent iterations of the MISCP optimization algorithm. Some embodiments of the invention are based on the recognition that adding one or more additional linear and / or smooth nonlinear inequality constraints to the MINLP constraints 633 can result in MICP subproblems 650 that are computationally cheaper to solve in each iteration of the MISCP optimization algorithm.

[0119]

number

[0120]

number

[0121] In some embodiments of the invention, if a particular set of values ​​of one or more integer and / or continuous variables is detected as infeasible for the MINLP 630, then each subsequent iteration of the MISCP optimization algorithm may include one or more linear and / or smooth nonlinear inequality constraints to avoid revisiting the same set of values ​​of the one or more integer and / or continuous variables for the MINLP 630.

[0122] Some embodiments of the present invention are based on a branch-and-bound method to solve the MICP subproblem at each iteration of the MISCP optimization algorithm, and adding one or more inequality constraints may result in a significantly smaller branch-and-bound search tree and thus significantly reduce the computational cost of solving the MICP subproblem.

[0123]

number

[0124]

number

[0125] FIG. 11A shows a schematic diagram of an example binary control variable search tree representing a nested tree of search spaces for integer feasible solutions of the MICP subproblem in the MISCP optimization algorithm, according to some embodiments. FIG. 11A shows a schematic diagram of a branch-and-bound method used to realize a MINMPC controller in some embodiments by showing a binary search tree 1100 at a particular iteration of the mixed integer optimization algorithm. The main idea of ​​the branch-and-bound method (B&B) is to sequentially create partitions of the original MICP subproblem and try to solve those partitions, each partition corresponding to a particular region of the discrete control variable search space. In some embodiments, the branch-and-bound method selects a partition or node and selects discrete control variables that branch this partition into smaller partitions or search spaces, resulting in a nested tree of partitions or search spaces.

[0126] For example, partition P1 1101 represents a discrete search region that can be split or branched into two smaller partitions or regions P2 1102 and P3 1103, i.e., a first region and a second region nested in a common region. The first and second regions are disjoint, i.e., their intersection is empty P2∩P3=φ (1107), but together they form the original partition or region P1, i.e., after branching, the union P2∪P3=P1 holds (1106). A branch-and-bound method then solves an integer relaxed convex problem for both the first and second partitions or regions of the search space, resulting in two solutions (local optima) that can be compared against each other and against the currently known upper bound on the optimal objective value. The first and / or second partitions or regions can be pruned if their performance metrics are less optimal than the currently known upper bound on the optimal objective value of the MICP subproblem. If the first region, the second region, or both regions yield a feasible discrete solution to the MICP subproblem, the upper bound can be updated. The branch-and-bound method then continues by selecting the remaining regions in the current nested tree of regions for further partitioning.

[0127] Although solving each partition may still be difficult, it is quite efficient to obtain a local lower bound on the optimal objective value by solving a local relaxation of the mixed integer program or by using duality. If the solution for the MICP subproblem happens to obtain an integer feasible solution while solving the local relaxation, the branch and bound method can then use it to obtain a global upper bound on the mixed integer control solution of the original MICP subproblem in the MISCP optimization algorithm to find a feasible and (locally) optimal solution for the MINLP. This helps to avoid solving or branching on certain partitions that have already been created, i.e., these partitions or nodes can be pruned. A general algorithmic concept of such a partition can be represented as a binary search tree 1100, including a root node, e.g., P1 1101, at the top of the tree, and leaf nodes, e.g., P4 1104 and P5 1105, at the bottom of the tree. In addition, nodes P21102 and P31103 are typically referred to as direct children of node P11101, and node P11101 is referred to as the parent of nodes P21102 and P31103. Similarly, nodes P41104 and P51105 are children of their parent node P21102.

[0128] FIG. 11B illustrates a block diagram of a branch-and-bound mixed integer optimization algorithm for searching an integer-feasible optimal solution of a MICP subproblem 320 based on the nested tree of the search space and the corresponding lower / upper bounds, according to some embodiments. The block diagram of the branch-and-bound mixed integer optimization algorithm illustrated in FIG. 11B can be used to realize a MINMPC controller, in some embodiments. The branch-and-bound method initializes 1110 the branch-and-search tree information of a mixed integer convex program (MICP) at the current control time step based on MICP data 1165 consisting of matrices and vectors. This initialization can further use the branch-and-search tree information and MICP solution information 1160 from previous iterations to generate a warm-start initialization 1110 for the current control time step. The main goal of the optimization algorithm is to construct upper and lower bounds for the objective value of the MICP subproblem solution. In step 1111, if the gap between the upper and lower bounds is smaller than a certain tolerance, a mixed integer optimal control solution 1155 is found.

[0129] In step 1111, as long as the gap between the lower and upper bounds is greater than a certain tolerance and the optimization algorithm has not yet reached its maximum run time, the branch and bound method continues the iterative search for a mixed integer optimal control solution 1155 of the MICP subproblem. Each iteration of the branch and bound method begins by selecting 1115 the next node in the tree, which corresponds to the next region or partition of the integer variable search space with possible variable fixes based on a pre-solution branching approach. After node selection, the corresponding integer relaxed convex problem is solved 1120 with possible variable fixes based on a post-solution branching approach.

[0130] If the integer relaxed convex problem has a feasible solution, then the resulting relaxed control solution provides a lower bound on the objective value for that particular region or partition of the integer variable search space. If, in step 1121, it is determined that the objective value is greater than the currently known upper bound on the objective value of the optimal mixed integer control solution of the MICP subproblem, then the selected node is pruned or removed from the branching tree (1140). However, if, in step 1121, it is determined that the objective value is lower than the currently known upper bound and the relaxed control solution is integer feasible (1125), then the currently known upper bound and the corresponding mixed integer control solution estimate are updated in step 1130.

[0131] If the integer relaxed convex problem has a feasible solution and the objective is lower than the currently known upper bound (1121), but the relaxed control solution is not yet integer feasible (1125), then the global lower bound of the objective can be updated to the minimum of the objective values ​​of the remaining leaf nodes in the branching tree (1135), and the selected node is pruned from the tree (1140). Additionally, starting from the current node, discrete variables with fractional values ​​are selected for branching (1145) according to a particular branching strategy to create a resulting preliminary MICP subproblem that corresponds to a small region or partition of the discrete search space of the original MICP subproblem and add it as a child of that node in the branching tree (1150).

[0132] The key step in the branch-and-bound method is how to form the partition, i.e., which nodes to select (1115) and which discrete variables to select for branching (1145). Some embodiments are based on branching one of the binary control variables with fractional values ​​in the integer relaxed convex problem solution. For example, a particular binary control variable u i,k If ∈ {0,1} has fractional values ​​as part of the integer relaxed convex problem solution, some embodiments may implement the equality constraints u i,k = 0 in one subproblem, and the equality constraint u i,k= 1 to the other subproblem. Some embodiments are based on a reliability branching strategy of variable selection (1145), which aims to predict future branching behavior based on information from previous branching decisions.

[0133] Some embodiments are based on a branch-and-bound method using a depth-first node selection strategy that can be implemented using a last-in-first-out (LIFO) buffer. The next node to be solved is selected as one of the children of the current node, and the process is repeated until the node is pruned, i.e., until it becomes infeasible, optimal, or dominated by a currently known upper bound, followed by a backtracking procedure. Instead, some embodiments are based on a branch-and-bound method using a best-first strategy that selects the node that currently has the lowest local lower bound. Some embodiments employ a combination of a depth-first node selection strategy and a best-first node selection strategy, where a depth-first node selection strategy is used until an integer feasible control solution is found, and then a best-first node selection strategy is used in subsequent iterations of the branch-and-bound based optimization algorithm. The latter implementation is motivated by the goal of finding an integer feasible control solution early at the start of the branch-and-bound procedure (depth-first) to allow early pruning, followed by a more greedy search for better feasible solutions (best-first).

[0134] The branch-and-bound method continues iteratively until one or more of the following conditions are met:

[0135] The processor's maximum execution time is reached.

[0136] All nodes in the branch search tree are pruned so that no new nodes can be selected for solving or branching the convex relaxation.

[0137] The optimality gap between the global upper and lower bounds of the objective of the MICP subproblem solution is smaller than the tolerance value.

[0138] 12A illustrates a schematic diagram of a vehicle 1201 including a predictive controller 1202 employing principles of some embodiments. As used herein, the vehicle 1201 may be any type of wheeled vehicle, such as a car, a bus, or a rover. The vehicle 1201 may also be an autonomous or semi-autonomous vehicle. For example, some embodiments control the motion of the vehicle 1201. Examples of motion include lateral motion of the vehicle controlled by a steering system 1203 of the vehicle 1201. In one embodiment, the steering system 1203 is controlled by the controller 1202. Additionally or alternatively, the steering system 1203 may be controlled by a driver of the vehicle 1201.

[0139] The vehicle may also include an engine 1206, which may be controlled by the controller 1202 or other components of the vehicle 1201. The vehicle may also include one or more sensors 1204 for sensing the surrounding environment. Examples of the sensors 1204 include distance range finders, radar, lidar, and cameras. The vehicle 1201 may also include one or more sensors 1205 for sensing its current momentum and internal state. Examples of the sensors 1205 include a global positioning system (GPS), an accelerometer, an inertial measurement unit, a gyroscope, a shaft rotation sensor, a torque sensor, a deflection sensor, a pressure sensor, and a flow sensor. The sensors provide information to the controller 1202. The vehicle may be equipped with a transceiver 1207 that enables communication functions of the controller 1202 via wired or wireless communication channels.

[0140] 12B illustrates a schematic diagram of the interaction between a controller 1202, e.g., a mixed integer nonlinear model predictive controller (MINMPC), and other controllers 1220 of the vehicle 1201, according to some embodiments. For example, in some embodiments, the controllers 1220 of the vehicle 1201 are a steering controller 1225 and a brake / throttle controller 1230 that control the rotation and acceleration of the vehicle 1220, respectively. In such a case, the predictive controller 1202 outputs control inputs to the controllers 1225 and 1230 to control the state of the vehicle 1201. The controller 1220 may also include a high-level controller, such as a lane-keeping assist controller 1235, that further processes the control inputs of the predictive controller 1202. In both cases, the controller 1220 uses the output of the predictive controller 1202 to control at least one actuator of the vehicle 1201, such as the steering wheel and / or brakes of the vehicle 1201, to control the motion of the vehicle 1201. Further, the predictive controller 1202 determines inputs to the vehicle 1201 based on the mixed integer optimal control solution, where the inputs to the vehicle 1201 include one or a combination of the acceleration of the vehicle 1201, the engine torque, the brake torque, and the steering angle of the vehicle 1201, as well as discrete optimization variables for modeling one or a combination of discrete control decisions, switching, gear shifting, and obstacle avoidance constraints in the system dynamics.

[0141] 12C shows a schematic diagram of a path and / or motion planning method for a controlled vehicle employing the principles of some embodiments. FIG. 12C shows a schematic diagram of an autonomous or semi-autonomous controlled vehicle 1250 that can use embodiments of the present disclosure to calculate a dynamically feasible and often optimal trajectory 1255. The generated trajectory aims to keep the vehicle within a particular road boundary 1252 and to avoid other controlled and / or uncontrolled vehicles, i.e., these vehicles are obstacles 1251 to this particular controlled vehicle 1250. In some embodiments, each of the obstacles 1251 can be represented by one or more inequality constraints in a time or space formulation of a constrained mixed integer nonlinear programming problem, including one or more additional discrete variables for each obstacle. For example, based on an embodiment configured to implement a mixed integer nonlinear model predictive controller, an autonomous or semi-autonomous controlled vehicle 1250 can make discrete decisions in real time, such as to overtake another vehicle on the left or right side, or alternatively, to stay behind another vehicle in the current lane of road 1252, while additionally making continuous decisions in real time, such as speed, acceleration, or steering inputs, to control the motion of the vehicle 1250.

[0142] FIG. 12D illustrates an example traffic scene for a single-vehicle or multi-vehicle decision-making module according to some embodiments. FIG. 12D depicts a scenario with one or more vehicles under control, referred to as ego-vehicle 1271, traffic consisting of other vehicles shown similar to 1272, lanes marked 1273, e.g., L6, stop lines marked 1274, e.g., S1, and intersections marked 1275, e.g., I3. For a vehicle at position 1261 with final destination 1262, the routing module provides a sequence of roads indicated by arrow 1263 and a sequence of direction changes indicated by arrow 1264. However, it should be noted that the road sequence 1263 and the direction change sequence 1264 do not in themselves specify the vehicle's trajectory or path. There are several discrete decisions to be made, such as which lane the vehicle should travel in, whether the vehicle should change lanes or stay in the current lane, whether the vehicle should start to slow down and stop at a stop line, whether the vehicle is allowed to cross an intersection, etc. In addition, there are several continuous decisions to be made, such as the timed sequence of positions and orientations the vehicle should achieve in its journey from its initial point to its destination. These decisions are highly dependent on the current traffic at the moment the vehicle reaches the corresponding position, which is generally unknown to the routing module due to the uncertainty of traffic motion and the uncertainty of the moment the vehicle will reach the position. In some embodiments of the present disclosure, the motion plan can be calculated for one or more controlled ego-vehicles 1271 by solving one or more connected mixed integer nonlinear programming problems, possibly with communication enabling vehicle-to-vehicle (V2V) and / or smart infrastructure system-to-vehicle (V2X) cooperation.

[0143] 13A and 13B are schematic diagrams of a spacecraft mixed integer nonlinear model predictive control problem formulation using the principles of some embodiments of the present disclosure. FIG. 13A and 13B show a spacecraft 1302 with multiple actuators, such as thrusters 1350 and momentum exchangers 1351. Examples of types of momentum exchangers include reaction wheels (RW) and gyroscopes. The spacecraft 1302 is a vehicle, vessel, or machine designed to fly in space, whose operations change quantities such as the position, velocity, attitude, or orientation of the spacecraft 1302 in response to commands sent to the actuators. When commanded, the actuators provide forces to the spacecraft 1302 that increase or decrease its velocity, thereby causing the spacecraft 1302 to translate its position, and when commanded, the actuators also provide torques to the spacecraft 1302, causing the spacecraft 1302 to rotate, thereby changing its attitude or orientation. As used herein, the motion of spacecraft 1302 is determined by the operation of actuators that determine the movements of spacecraft 1302 that change such quantities.

[0144] The spacecraft 1302 flies in space along an open or closed orbital path 1360 around, between, or near one or more gravitational bodies, such as the Earth 1361, the Moon, and / or other celestial planets, stars, asteroids, comets, etc. Typically, a desired or target position 1365 along the orbital path is given. A reference frame 1370 is attached to the desired location, and the origin of the frame, i.e., the all-zero coordinate in that reference frame, is always the coordinate of the desired location.

[0145] The spacecraft 1302 is subject to various disturbance forces 1314. These disturbance forces may include forces that were not considered in determining the orbital path of the spacecraft 1302. These disturbance forces act on the spacecraft 1302 to move the spacecraft 1302 away from a desired position on the orbit. These forces include, but are not limited to, gravitational attraction, radiation pressure, atmospheric drag, non-spherical central bodies, and propellant leakage. Thus, the spacecraft 1302 may be at a distance 1367 away from the target position.

[0146] Due to disturbance forces, it is not always possible to maintain spacecraft 1302 at a desired location along its orbit. Therefore, instead, it is desired that spacecraft 1302 remain within a window 1366 with specified dimensions 1364 around the desired location. To do so, spacecraft 1302 is controlled to move along any path 1380 that falls within the desired target window. In this example, window 1366 has a rectangular shape, although the shape of the window may vary from embodiment to embodiment.

[0147] Spacecraft 1302 is often also required to maintain a desired orientation. For example, a reference frame 1374 fixed to the spacecraft needs to be aligned with a desired reference frame, such as an inertial reference frame 1371 fixed relative to a distant star 1372, or a reference frame 1373 oriented in such a way that it always points toward the Earth. However, depending on the shape of the spacecraft 1302, different disturbance forces 1314 may act non-uniformly on the spacecraft 1302, which may generate disturbance torques and rotate the spacecraft 1302 away from its desired orientation. To compensate for the disturbance torques, a momentum exchange device 1351, such as a reaction wheel, is used to absorb the disturbance torques, thereby allowing the spacecraft to maintain its desired orientation.

[0148] To prevent the momentum exchange devices from becoming saturated and thereby losing their ability to compensate for disturbance torques, their stored momentum must be unloaded, for example by slowing down the spin rate of the reaction wheels. Unloading the momentum exchange devices would impose undesirable torques on the spacecraft 1302. Such undesirable torques are also compensated for by the thrusters.

[0149] In some embodiments of the invention, one or more nonlinear equations are used to describe the dynamics and / or constraints of the controlled system, and the MINMPC controller determines inputs to the spacecraft 1302 based on a mixed integer optimal control solution, which inputs to the spacecraft 1302 actuate one or a combination of thrusters and momentum exchange devices, and discrete optimization variables are used to model one or a combination of discrete control decisions, switches in the system dynamics, integer values ​​of thruster commands, and obstacle avoidance constraints.

[0150] In some embodiments, the spacecraft 1302 can be modeled as a hybrid system, and the commands sent to the actuators are calculated using a predictive controller, such as a mixed integer nonlinear model predictive controller (MINMPC). For example, in some embodiments, the commands sent to the thrusters 1350 can only take on a discrete set of values, and thus result in a set of binary or integer control input variables for each stage in the mixed integer control horizon.

[0151] In some embodiments of the invention, the predictive controller is designed to keep the spacecraft 1302 outside a particular zone 1385 of particular dimensions close to the desired location along the orbit. The latter zone can be fixed in time or can vary in time and is often referred to as the exclusion zone 1385, for which corresponding logical inequality constraints can be modeled with an additional set of binary or integer control input variables for each stage in the mixed-integer control horizon. In this example, the exclusion zone 1385 is rectangular and positioned at the corners of the desired window 1366, although the shape and location of the exclusion zone within the desired target window may vary in different embodiments.

[0152] 14A shows a schematic diagram of a vapor compression system 1400 controlled by a controller 1460, according to some embodiments. The controller 1460 includes a predictive controller, such as a controller implementing mixed integer nonlinear model predictive control (MINMPC). Components of the vapor compression system (VCS) 1400 may include an indoor unit heat exchanger 1420 located in an indoor space or zone 1450, an outdoor unit heat exchanger 1430 located in the ambient environment, a compressor 1410, and an expansion valve 1440. A heat load 1415 acts on the indoor space or zone 1450.

[0153] Additionally, the VCS 1400 may include a check valve 1455 that is used to direct high pressure refrigerant exiting the compressor to either the outdoor unit heat exchanger or the indoor unit heat exchanger, and direct low pressure refrigerant returning from either the indoor unit heat exchanger or the outdoor unit heat exchanger to the inlet of the compressor. When high pressure refrigerant is directed to the outdoor unit heat exchanger, the outdoor unit heat exchanger acts as a condenser and the indoor unit acts as an evaporator, and the system rejects heat from the zone to the ambient environment, which is referred to as a "cooling mode" of operation. Conversely, when high pressure refrigerant is directed to the indoor unit heat exchanger, the indoor unit heat exchanger acts as a condenser and the outdoor unit heat exchanger acts as an evaporator, extracting heat from the ambient environment and pumping this heat into the zone, which is referred to as a "heating mode" of operation.

[0154] FIG. 14B illustrates an example of a configuration of signals, sensors, and controllers used in the VCS 1400, according to some embodiments. The controller 1460 reads information from sensors 1470 configured to measure various temperatures, pressures, flows, and other information related to the operation of the system, including measurable disturbances such as ambient air temperature. The controller 1460 can be provided with set points 1466 that represent desired values ​​of measured signals of the process, such as desired zone temperatures. Set point information can be obtained from a thermostat, a wireless remote control, or an internal memory or storage medium. The controller then calculates control inputs such that a number of measured outputs are driven to those set points. These control inputs can include indoor unit fan speed 1480, outdoor unit fan speed 1481, compressor rotation speed 1482, expansion valve position 1483, and backflow valve position 1484. In this manner, the controller controls the operation of the vapor compression system such that set point values ​​are achieved in the presence of disturbances 1468, such as heat loads acting on the system.

[0155] In some embodiments, the VCS 1400 can be modeled as a hybrid nonlinear system, and the commands sent to the actuators are calculated using a predictive controller, such as a mixed integer nonlinear model predictive controller (MINMPC). For example, in some embodiments, the commands sent to valves and / or fans can only take on a discrete set of values, and thus result in a set of binary or integer control input variables for each stage in the mixed integer control horizon.

[0156] In some embodiments, the predictive controller determines inputs to the vapor compression system based on a mixed integer optimal control solution, where the inputs to the vapor compression system include one or a combination of indoor unit fan speed, outdoor unit fan speed, compressor rotational speed, expansion valve position, and backflow valve position, and uses discrete optimization variables to model one or a combination of integer values ​​for discrete control decisions, switching in the system dynamics, and commands sent to the valves and / or fans.

[0157] In some embodiments, the dynamic behavior of the VCS 1400 may change rapidly or even switch at specific instants depending on the current state of the system and the current control input values. The resulting hybrid VCS 1400 with switching dynamics can be modeled with an additional set of binary or integer control input variables for each stage in the mixed-integer control horizon.

[0158] 15 illustrates a method 1500 for controlling a system, according to an example embodiment. In step 1501, the method includes accepting a feedback signal that includes a measurement of a state of the system. In step 1503, the method includes solving a mixed integer nonlinear optimal control problem using a mixed integer sequential convex programming (MISCP) optimization algorithm that searches for a feasible and (locally) optimal solution of a mixed integer nonlinear programming (MINLP) problem. In step 1505, the system is controlled based on the control signal to change the state of the system.

[0159] The above-described embodiments of the present disclosure may be implemented in any of numerous ways. For example, the embodiments may be implemented using hardware, software, or a combination thereof. If implemented in software, the software code may be executed on any suitable processor or collection of processors, whether provided on a single computer or distributed across multiple computers. Such a processor may be implemented as an integrated circuit with one or more processors within an integrated circuit component. However, the processor may be implemented using any suitable form of circuitry.

[0160] Also, the various methods or processes outlined herein may be coded as software executable on one or more processors using any one of a variety of operating systems or platforms. In addition, such software may be written using any of a number of suitable programming languages ​​and / or programming or scripting tools, and compiled as executable machine language code or intermediate code that runs on a framework or virtual machine. Typically, the functionality of the program modules may be combined or distributed as desired in various embodiments.

[0161] Also, embodiments of the present disclosure may be embodied as methods, for which examples are provided. The acts performed as part of the method may be ordered in any suitable manner. Thus, embodiments may be constructed in which acts are performed in an order different from that illustrated, which may include performing simultaneously some acts shown as sequential acts in the exemplary embodiments.

[0162] Although the disclosure has been described by way of examples of preferred embodiments, it will be understood that various other adaptations and modifications can be made within the scope of the disclosure. Therefore, it is the object of the appended claims to cover all such variations and modifications that come within the true spirit and scope of the disclosure.

Claims

1. 1. A predictive feedback controller for controlling a hybrid dynamic system having nonlinear dynamics and continuous and discrete elements of motion, comprising: at least one processor; and a memory having instructions stored thereon, the instructions, when executed by the at least one processor, causing the predictive feedback controller to: accepting a feedback signal including measurements indicative of a current state of the hybrid dynamic system, including one or a combination of a current state of the predictive controller, a current state of one or more actuators of the hybrid dynamic system, and a current state of an output of the hybrid dynamic system; and formulating a mixed integer nonlinear programming (MINLP) problem that optimizes an objective function subject to one or more constraints with a solution that indicates control commands for modifying the current state of the hybrid dynamic system according to a control objective, the constraints including equality constraints, inequality constraints, or both, the constraints and the control objective of the MINLP problem including one or more nonlinear functions of continuous optimization variables that represent the continuous elements of the hybrid dynamic system's motion and one or more linear functions of integer optimization variables that represent the discrete elements of the hybrid dynamic system's motion, the MINLP problem being formulated in a separable format that ensures that the discrete elements of the motion exist only in the linear functions of the MINLP problem, and further wherein the instructions, when executed by the at least one processor, cause the predictive feedback controller to: To solve the MINLP problem over multiple iterations of a sequential convexification-based optimization procedure until a termination condition is met, the predictive feedback controller is configured to perform a partial convexification of a portion of the solution space that includes a current solution guess, the partial convexification generating a convex approximation of the nonlinear function of the MINLP without approximating the linear function of the MINLP to generate a partially convexified MINLP, and further, the instructions, when executed by the at least one processor, cause the predictive feedback controller to update the current solution guess by solving a mixed integer convex programming (MICP) formulation of the partially convexified MINLP problem; A predictive feedback controller that submits the control commands generated according to the solution of the MINLP problem to the hybrid dynamic system, thereby causing a change in the current state of the hybrid dynamic system.

2. 2. The predictive feedback controller of claim 1, wherein the hybrid dynamic system includes a device controlled by an operational input signal representing the control command, the input signal specifying values ​​of the continuous elements of the operation of the hybrid dynamic system and values ​​of the discrete elements of the operation of the hybrid dynamic system, the continuous elements including one or a combination of voltage, pressure, force, torque, steering angle, speed and temperature, and the discrete elements including one or a combination of energy levels, quantized valve inputs, gear shifting, on / off actuation, lane selection, and obstacle avoidance decision variables.

3. The processor, 2. The predictive feedback controller of claim 1, configured to convert the MINLP problem from the original format to the separable format by adding one or a combination of an additional integer optimization variable and an additional continuous optimization variable to the original format, thereby ensuring that the discrete elements of the motion exist only in the linear functions of the MINLP in a separable format.

4. To perform the partial convexification, the processor Identifying whether the objective function of the MINLP problem or a function in the constraints is a function of the integer optimization variables or the continuous optimization variables; if the function is a smooth nonlinear function of the continuous optimization variables, computing a convex approximation of the smooth nonlinear function in a local neighborhood of the current solution guess within the solution space of the MINLP problem; 2. The predictive feedback controller of claim 1, configured to avoid the convex approximation by preserving the integer optimization variables when the function is a linear function of the integer optimization variables and is in a non-convex portion of the solution space of the MINLP.

5. 5. The predictive feedback controller of claim 4, wherein the convex approximation is computed by a locally constrained linearization or linear-quadratic objective approximation of the smooth nonlinear function in a local neighborhood of the current solution guess in the solution space of the MINLP problem by evaluating one or more first and / or higher order directional derivatives of the smooth nonlinear function using symbolic, numerical or algorithmic differentiation.

6. To update the current solution guess for the iteration, the predictive feedback controller: solving the MICP problem of the partially convexified MINLP to compute search directions for the continuous optimization variables and search directions for the integer optimization variables; updating current values ​​of the integer optimization variables toward the search direction of the integer optimization variables satisfying a constraint for each of the integer optimization variables being integer-valued; 2. The predictive feedback controller of claim 1 configured to update current values ​​of the continuous optimization variables toward the search direction of the continuous optimization variables with step-size values ​​governed by feasibility and optimality conditions of the MINLP problem.

7. 7. The predictive feedback controller of claim 6, wherein the step size value is between 0 and 1 and is selected for the iterations based on a merit function that balances optimality and feasibility of the solution to the MINLP problem, the value of the merit function decreasing between at least two iterations of the successive convexification-based optimization procedure.

8. The step size value is selected based on an iterative line search procedure starting from an initial step size value of one, the predictive feedback controller performing one or more iterations of the iterative line search procedure to select the step size value, each iteration comprising: testing whether a reduction condition is satisfied for a merit function that balances optimality and feasibility of the solution to the MINLP problem; If a sufficient decrease condition is met, terminate the iterative line search procedure and use the step size value to update the current value of the continuous optimization variable towards the search direction of the continuous optimization variable; otherwise, The predictive feedback controller of claim 7 configured to decrease the step size value and continue the iterative line search procedure.

9. 8. The predictive feedback controller of claim 7, wherein the merit function is a combination of the objective function and one or more penalty functions that are applied to violations of each of the equality constraints and violations of each of the inequality constraints in the MINLP problem.

10. 7. The predictive feedback controller of claim 6, wherein the MICP problem of the partially convexified MINLP includes one or more trust region inequality constraints to ensure that the search direction of the continuous optimization variables is below a trust region radius value.

11. To update the current solution guess for the iteration, the processor: Calculate a ratio of actual reduction to predicted reduction for a merit function value that balances optimality and feasibility of the MICP solution of the partially convexified MINLP problem; 11. The predictive feedback controller of claim 10, configured to update current values ​​of the integer and continuous optimization variables toward the search direction of the integer and continuous optimization variables if a ratio of the actual reduction to a predicted reduction in the value of the merit function exceeds a threshold.

12. To update the trust region radius value based on the solution of the MICP problem of the partially convexified MINLP, the processor: if the ratio of the actual reduction to the predicted reduction in the value of the merit function or the search direction of the continuous optimization variables falls below a threshold, decreasing the trust region radius value to ensure that the search direction of the continuous optimization variables is reduced in one or more of the next iterations, otherwise The predictive feedback controller of claim 11 , further comprising: increasing the trust region radius value to ensure that the search direction of the continuous optimization variables is increased in one or more subsequent iterations.

13. The predictive feedback controller of claim 1 , wherein a branch and bound (B&B) optimization method is used to compute a globally optimal solution to the MICP problem of the partially convexified MINLP.

14. 14. The predictive feedback controller of claim 13, wherein the MICP problem is a mixed integer linear programming (MILP) subproblem, a mixed integer quadratic programming (MIQP) subproblem, a mixed integer quadratically constrained quadratic programming (MIQCQP) subproblem, or a mixed integer quadratic cone programming (MISOCP) subproblem.

15. The current values ​​of the integer optimization variables are fixed after one or more iterations of the sequential convexification-based optimization procedure, and to update the current solution guesses in one or more iterations of the sequential convexification-based optimization procedure, the predictive feedback controller: solving a convex programming (CP) problem for the partially convexified MINLP given the fixed current values ​​of the integer optimization variables to compute search directions for the continuous optimization variables; 7. The predictive feedback controller of claim 6, configured to update current values ​​of the continuous optimization variables toward the search direction of the continuous optimization variables with step-size values ​​governed by feasibility and optimality conditions of the MINLP problem.

16. 16. The predictive feedback controller of claim 15, wherein the CP problem is a convex linear programming (LP) subproblem, a convex quadratic programming (QP) subproblem, a convex quadratically constrained quadratic programming (QCQP) subproblem, or a convex quadratic cone programming (SOCP) subproblem.

17. 2. The predictive feedback controller of claim 1, further comprising: a homotopy-based penalty method for penalizing changes in values ​​of one or more of the integer optimization variables based on a positive weight value between at least two iterations of the sequential convexification-based optimization procedure.

18. 20. The predictive feedback controller of claim 17, wherein the homotopy-based penalty method adds one or more penalty terms in the objective function of the MICP problem of the partially convexified MINLP to accelerate the convergence of the successive convexification-based optimization procedure to a feasible and / or optimal solution of the MINLP problem.

19. 18. The predictive feedback controller of claim 17, wherein the homotopy-based penalty method adds one or a combination of multiple linear and / or smooth nonlinear inequality constraints to the inequality constraints of the MICP problem of the partially convexified MINLP to accelerate the convergence of the successive convexification-based optimization procedure to a feasible and / or optimal solution of the MINLP problem.

20. 2. The predictive feedback controller of claim 1, wherein the at least one processor is at least one parallel processor, and the solution to the MICP problem of the partially convexified MINLP can be found by searching in parallel for globally optimal solutions to the MICP problem in different regions of the solution space.

21. The control commands generated by the predictive feedback controller specify a target state of the hybrid dynamic system, the hybrid dynamic system having:

2. The hybrid dynamic system including the predictive feedback controller of claim 1, comprising a tracking controller configured to generate one or more control inputs to the actuators of the hybrid dynamic system to reduce an error between the current state of the hybrid dynamic system and the target state of the hybrid dynamic system.

22. The hybrid dynamic system of claim 21 , wherein the tracking controller is a PID controller or a model predictive controller (MPC).

23. 2. The predictive feedback controller of claim 1, wherein the predictive feedback controller is realized using a mixed integer nonlinear model predictive control (MINMPC), the MINMPC calculating the control signals based on current states and control commands of the system, and the MINMPC calculating a control solution comprising a sequence of future optimal discrete and continuous control inputs over a prediction time horizon of the hybrid dynamic system by solving a constrained mixed integer nonlinear optimization problem at each control time step.

24. 24. The predictive feedback controller of claim 23, wherein a solution of the constrained mixed integer nonlinear optimization problem at one control time step is used as an initial solution guess for the successive convexification based optimization procedure to calculate a solution of the constrained mixed integer nonlinear optimization problem at a next control time step.

25. 2. The predictive feedback controller of claim 1, wherein the system is a vehicle, and the predictive feedback controller determines inputs to the vehicle based on the MINLP solution, the inputs to the vehicle including one or a combination of acceleration of the vehicle, engine torque, brake torque, and steering angle of the vehicle, and the discrete optimization variables are used to model one or a combination of discrete control decisions, switching, gear shifting, and obstacle avoidance constraints in system dynamics.

26. 2. The predictive feedback controller of claim 1, wherein the system is a spacecraft, and the predictive feedback controller determines inputs to the spacecraft based on the MINLP solution, the inputs to the spacecraft actuating one or a combination of thrusters and momentum exchange devices, and the discrete optimization variables are used to model one or a combination of discrete control decisions, switches in the system dynamics, integer values ​​for thruster commands, and obstacle avoidance constraints.

27. 2. The predictive feedback controller of claim 1, wherein the system is a vapor compression system, and the predictive feedback controller determines inputs to the vapor compression system based on the MINLP solution, the inputs to the vapor compression system including one or a combination of indoor unit fan speed, outdoor unit fan speed, compressor rotational speed, expansion valve position, and backflow valve position, and the discrete optimization variables are used to model one or a combination of discrete control decisions, switching in system dynamics, and integer values ​​for commands sent to the valves and / or the fans.

Citation Information

Patent Citations

  • Process operation planning system

    JP2003295906A

  • Control performance calculation method and control performance calculation device for plant

    JP2014078146A

  • Problem solution device, method, and program

    JP2020144529A

  • Process optimization using mixed integer nonlinear programming

    US20160370772A1