9. High-level Interface

The FORCESPRO high-level interface gives optimization users a familiar easy-to-use way to define an optimization problem. The interface also gives the advanced user full flexibility when importing external C-coded functions to evaluate the quantities involved in the optimization problem.

This interface is provided with Variant L and partially with Variant M of FORCESPRO.

Important

Starting with FORCESPRO 1.8.0, the solver generated from the high-level interface supports nonlinear and convex decision making problems with integer variables.

Note

The high-level Python interface expects 0-based indices in the model formulation, such as for the indices in lbidx, ubidx, hlidx, huidx, xinitidx and xfinalidx, as is usual in Python programs. Note that this is contrary to the low-level interface, which requires 1-based indices for these fields.

9.1. Supported problems

9.1.1. Canonical problem for discrete-time dynamics

The FORCESPRO NLP solver solves (potentially) non-convex, finite-time nonlinear optimal control problems with horizon length \(N\) of the form:

\begin{align*} \text{minimize} \quad & \sum_{k=1}^{N} f_k(z_k,{\color{red} p_{\color{red} k}}) && \quad \text{(separable objective)} \\ \text{subject to} \quad & z_1(\mathcal{I}) = {\color{red} z_{\mathrm{init}}} && \quad \text{(initial equality)} \\ & E_k z_{k+1} = c_k(z_k,{\color{red} p_{\color{red} k}}) && \quad \text{(inter-stage equality)} \\ & \underline{h}_k \leq h_k(z_k,{\color{red} p_{\color{red} k}}) \leq \bar{h}_k && \quad \text{(nonlinear constraints)} \\ & z_N(\mathcal{N}) = {\color{red} z_{\mathrm{final}}} && \quad \text{(final equality)} \\ & \underline{z}_k \leq z_k \leq \bar{z}_k && \quad \text{(lower-upper bounds)} \\ & F_kz_k \in \left[\underline{z}_k, \bar{z}_k\right]\cap\mathbb{Z} && \quad \text{(integer variables)} \end{align*}

for \(k = 1,\dots,N\), where \(z_k \in \mathbb{R}^{n_k}\) are the optimization variables, for example a collection of inputs, states or outputs in an MPC problem; \({\color{red} p_{\color{red} k}} \in \mathbb{R}^{l_k}\) are real-time data, which are not necessarily present in all problems; the functions \(f_k : \mathbb{R}^{n_k} \times \mathbb{R}^{l_k} \rightarrow \mathbb{R}\) are stage cost functions; the functions \(c_k : \mathbb{R}^{n_k} \times \mathbb{R}^{l_k} \rightarrow \mathbb{R}^{w_k}\) represents (potentially nonlinear) equality constraints, such as a state transition function; the matrices \(E_k\) are used to couple variables from the \((k+1)\)-th stage to those of stage \(k\) through the function \(c_k\); and the functions \(h_k : \mathbb{R}^{n_k} \times \mathbb{R}^{l_k} \rightarrow \mathbb{R}^{m_k}\) are used to express potentially nonlinear, non-convex inequality constraints. The index sets \(\mathcal{I}\) and \(\mathcal{N}\) are used to determine which variables are fixed to initial values \(\color{red} z_{\mathrm{init}}\) and final values \(\color{red} z_{\mathrm{final}}\), respectively; initial and final values can also be changed in real-time. Finally, lower and upper bounds \(\underline{z}_k < \bar{z}_k\) on \(z_k\) can be specified at every stage \(k\); the matrix \(F_k\) is a selection matrix that may select certain coordinates in vector \(z_k\) to be integer.

All real-time data is coloured in red. Additionally, when integer variables are modelled, they need to be declared as real-time parameters (see section Mixed-integer nonlinear solver).

To obtain a solver for this optimization problem using the FORCESPRO client, you need to define all functions involved \((f_k, c_k, h_k)\) along with the corresponding dimensions.

9.1.2. Continuous-time dynamics

Instead of having discrete-time dynamics as can be seen in Section 9.1.1, we also support using continuous-time dynamics of the form:

\[\dot{x} = f(x,u,{\color{red}p})\]

and then discretizing this equation by one of the standard integration methods. See Section 9.2.4.2 for more details.

9.1.3. Other variants

Not all elements in Section 9.1.1 have to be necessarily present. Possible variants include problems:

  • where all functions are fixed at code generation time and do not need extra real-time data \(\color{red} p\);

  • with no lower (upper) bounds for variable \(z_{k,i}\), then \(\underline{z}_i \equiv -\infty~(\bar{z}_i \equiv + \infty)\);

  • without nonlinear inequalities \(h\);

  • with \(N = 1\) (single stage problem), then the inter-stage equality can be omitted;

  • with final equality, then the final stage cost function \(f_N\) can be omitted;

  • that optimize over the initial value \(\color{red} z_{\mathrm{init}}\) and do not include the initial equality;

  • that optimize over the final value \(\color{red} z_{\mathrm{final}}\) and do not include the final equality;

  • mixed-integer nonlinear programs, where some variables are declared as integers (see section Mixed-integer nonlinear solver for more information about the MINLP solver).

9.1.4. Function evaluations

The FORCESPRO NLP solver requires external functions to evaluate:

  • the cost function terms \(f_k(z_k)\) and their gradients \(\nabla f_k(z_k)\),

  • the dynamics \(c_k(z_k)\) and their Jacobians \(\nabla c_k (z_k)\), and

  • the inequality constraints \(h_k(z_k)\) and their Jacobians \(\nabla h_k(z_k)\).

The FORCESPRO code generator supports the following ways of supplying these functions:

1. Automatic C-code generation of these functions from MATLAB using a supported automatic differentiation (AD) tool, such as CasADi. This happens automatically in the background, as long as the AD tool is found on your system. By doing so, the user does not need to adhere to any tool-specific syntax but can use standard MATLAB commands to define the necessary functions instead (which are then automatically converted to match the specifics of the chosen AD tool). This is the recommended way of getting started with FORCESPRO NLP. See Section 9.2 to learn how to use this approach.

2. C-functions (source files). These can be hand-coded, or generated by any automatic differentiation tool. See Section 9.5 for details on how to provide own function evaluations and derivatives to FORCESPRO.

9.2. Expressing the optimization problem in code

When solving nonlinear programs of the type in Section 9.1.1, FORCESPRO requires the functions \(f,c,h\) and their derivatives (Jacobians) to be evaluated in each iteration. There are two ways for accomplishing this: either implement all function evaluations in C by some other method (by hand or by another automatic differentiation tool), or use an AD tool integrated with FORCESPRO, such as the open-source package CasADi (see Automatic differentiation tool for a list of all supported tools). This is the easiest option to quickly get started with solving NLPs, and it generates efficient code. However, if you prefer other AD tools, see Section 9.5 to learn how to provide your own derivatives to FORCESPRO NLP solvers. This section will describe the CasADi-based approach in detail, using either the MATLAB or the Python client of FORCESPRO. Please note that even though both the MATLAB and the Python client are intended to behave largely identical, there are some differences between the two clients. For details, refer to Differences between the MATLAB and the Python client.

9.2.1. Model Initialization

9.2.1.1. Model Initialization in MATLAB

In the MATLAB high-level interface, the formulation of the optimization problem is given through a simple structure array. In the following, we will describe the problem in such an array named model. It is advisable to zero-initialize this variable at the beginning of your script such that no values set in previous iterations of your script interfere with this run:

model = {}

9.2.1.2. Model Initialization in Python

In the high-level Python interface, optimization problems are described through objects of different types, depending on the problem. The following classes are available:

  • SymbolicModel - Allows you to describe your optimization problem using regular Python functions. These functions will be evaluated symbolically by CasADi to create optimized C code. Note that this model is meant to be used for nonlinear models. If you wish to express a convex model symbolically, consider using the ConvexSymbolicModel or forcing generation of a nonconvex solver by setting the option forcenonconvex to True.

  • ExternalFunctionModel - Enables more flexibility in describing nonlinear problems by allowing any external function to be used as objective function and constraints. This requires C code or already compiled code (object files or shared libraries) from any language. The approach using external function evaluations for your objective function and constraints is described in External function evaluations in C, including the required call signature of the callback function.

  • ConvexSymbolicModel - FORCESPRO can generate different solvers for convex problems.

Whichever model you choose, it can be initialized with no arguments, or with a single argument denoting the number of stages \(N\) in the problem:

import forcespro.nlp
model = forcespro.nlp.SymbolicModel(50)

Note that most symbolic problem descriptions will also require the Numpy and CasADi packages, so it is a good idea to import them at the beginning:

import numpy as np
import casadi

9.2.2. Dimensions

In order to define the dimensions of the stage variables \(z_i\), the number of equality and inequality constraints and the number of real-time parameters use the following fields (properties) in the client:

model.N = 50;   % length of multistage problem
model.nvar = 6; % number of stage variables
model.neq = 4;  % number of equality constraints
model.nh = 2;   % number of nonlinear inequality constraints
model.npar = 0; % number of runtime parameters

If the dimensions vary for different stages use arrays of length \(N\). See Section 9.2.7.1 for an example.

9.2.3. Objective

The high-level interface allows you to define the objective function using a handle to a MATLAB or Python function that evaluates the objective. This function is called with the variables of one stage as its first argument, i.e. a vector of model.nvar entries. FORCESPRO will process the given function symbolically and generate the necessary C code to be included in the solver.

model.objective = @eval_obj; % handle to objective function

For instance, the function could be:

function f = eval_obj ( z )
    F = z(1);
    s = z(2);
    y = z(4);
    f = -100*y + 0.1*F^2 + 0.01* s^2;
end

If the cost function varies for different stages use a cell array of function handles of length \(N\) in MATLAB, or a list of function handles in Python. See Section 9.2.7.1 for an example.

Note

Python and MATLAB use different indexing bases. The first element of any variable has index 1 in MATLAB, whereas it is accessed at offset 0 in Python!

The objective evaluation function can optionally accept an additional argument p which serves as a run-time parameter. In order to be able to change the terms in the cost function during runtime, one can define the objective function as:

function f = eval_obj ( z, p )
    F = z(1);
    s = z(2);
    y = z(4);
    f = -100*y + p(1)*F^2 + p(2)* s^2;
end

The length of this additional parameter vector in each stage is given by model.npar.

9.2.4. Equalities

9.2.4.1. Discrete-time

For discrete-time dynamics, one can define a handle to a function evaluating \(c\) as shown below. The selection matrix \(E\) that determines which variables are affected by the inter-stage equality must also be filled. For performance reasons, it is recommended to order variables such that the selection matrix has the following structure:

model.eq = @eval_dynamics;      % handle to inter-stage function
model.E = [zeros(4,2), eye(4)]; % selection matrix

If the equality constraint function varies for different stages use a cell array (or list in Python) of function handles of length \(N-1\), and similarly for \(E_k\). See Section 9.2.7.1 for an example.

9.2.4.2. Continuous-time

For continuous-time dynamics, FORCESPRO requires you to describe the dynamics of the system in the following form:

\[\dot{x} = f(x,u,{\color{red}p})\]

where \(x\) are the states of the system, \(u\) are the inputs and \(\color{red} p\) a vector of parameters, e.g. the mass or inertia. The selection matrix \(E\) determines which components of the stage variable \(z_i\) are to be considered state \(x\) or input \(u\) in this representation.

For example, let’s assume that the system to be controlled has the dynamics:

\[\dot{x} = {\color{red} p_{\color{red} 1}} x_1 x_2 + {\color{red} p_{\color{red} 2}} u\]

In order to discretize the system for use with FORCESPRO we have to:

  1. Implement the continuous-time dynamics as a function:

function xdot = continuous_dynamics(x, u, p)
      xdot = p(1)*x(1)*x(2) + p(2)*u;
end

Note that in general the parameter vector p can be omitted if there are no parameters. You can also implement short functions as anonymous function handles:

continuous_dynamics_anonymous = @(x,u,p) p(1)*x(1)*x(2) + p(2)*u;

2. Tell FORCESPRO that you are using continuous-time dynamics by setting the continuous_dynamics field of the model to a function handle to one of the functions above:

model.continuous_dynamics = @continuous_dynamics;

or, if you are using anonymous functions:

model.continuous_dynamics = @continuous_dynamics_anonymous;

3. Use the selection matrix \(E\) to link the stage variables \(z_i\) with the states \(x\) and inputs \(u\) of the continuous dynamics function:

model.E = [zeros(2, 1), eye(2)]

Components of \(z_i\) are considered as state variables \(x\) according to the order prescribed by the selection matrix. If an entire \(k\)-th column of the selection matrix is zero, the \(k\)-th component of \(z_i\) is not governed by a dynamic equation and thus considered as input \(u\).

  1. Choose one of the integrator functions from the Integrators section (the default is ERK4):

codeoptions.nlp.integrator.type = 'ERK2';
codeoptions.nlp.integrator.Ts = 0.1;
codeoptions.nlp.integrator.nodes = 5;

where the integrator type is set using the type field of the options struct codeoptions.nlp.integrator. The field Ts determines the absolute time between two integration intervals, while nodes defines the number of intermediate integration nodes within that integration interval. In the example above, we use 5 steps to integrate for 0.1 seconds, i.e. each integration step covers an interval of 0.02 seconds.

9.2.5. Initial and final conditions

The indices affected by the initial and final conditions can be set as follows:

model.xinitidx = 3:6;   % indices affected by initial condition
model.xfinalidx = 5:6;  % indices affected by final condition

Note

Python and MATLAB use different indexing bases. The first variable in a stage has index 1 in MATLAB, whereas it is accessed at offset 0 in Python! Further note that Python’s range does not include the upper limit, thus:

list(range(2, 6)) == [2, 3, 4, 5]  # does not include upper limit

9.2.6. Inequalities

A function evaluating nonlinear inequalities can be provided in a similar way, for example:

function h = eval_const(z)
  x = z(3);
  y = z(4);
  h = [x^2 + y^2;
       (x+2)^2 + (y-2.5)^2 ];
end

Note

For Python installations with Numpy version 1.20 onwards it is advised to use CasADi arrays and CasADi functions (where available) for the implementation of the functions assigned to: model.objective, model.eq, model.ineq, model.continuous_dynamics for the problem formulation in order to ensure maximum compatibility between CasADi and the FORCESPRO Python client.

The simple bounds and the nonlinear inequality bounds can have +inf and -inf elements, but must be the same length as the field nvar and nh, respectively:

model.ineq = @eval_const;             % handle to nonlinear constraints
model.hu = [9, +inf];                 % upper bound for nonlinear constraints
model.hl = [1, 0.95^2];               % lower bound for nonlinear constraints
model.ub = [+5, +1, 0, 3, 2, +pi];    % simple upper bounds
model.lb = [-5, -1, -3, -inf,  0, 0]; % simple lower bounds

Note

While the FORCESPRO Python client does not require you to use numpy arrays, we encourage their use for vector- and matrix-valued properties of the model, as it simplifies calculations for the user. Therefore, any of the above properties can also be set to Numpy arrays instead of lists. If lists are given, these are converted to Numpy arrays internally.

If the constraints vary for different stages, use cell arrays of length \(N\) for any of the quantities defined above. See Varying dimensions, parameters, constraints, or functions section for an example.

Bounds model.lb, model.ub, model.hl and model.hu can be made parametric by leaving said fields empty and using the model.lbidx, model.ubidx, model.hlidx and model.huidx fields respectively to indicate on which variables/inequalities lower and upper bounds are present. The numerical values will then be expected at runtime. The runtime parameters will be created stage-wise for the above bounds and will have the names lb<n>, ub<n>, hl<n>, hu<n> where <n> will be the 1-based stage number they belong it (padded with enough 0 based on the largest stage) For example, to set parametric lower bounds on states 1 and 2, and parametric upper bounds on states 2 and 3, use:

% Lower bounds are parametric (indices not mentioned here are -inf)
model.lbidx = [1 2]';

% Upper bounds are parametric (indices not mentioned here are +inf)
model.ubidx = [2 3]';

% lb and ub have to be empty when using parametric bounds
model.lb = [];
model.ub = [];

and then specify the exact values at runtime through the fields lb01–lbN and ub01–ubN:

% Specify lower bounds
problem.lb01 = [0 0]';
problem.lb02 = [0 0]';
% ...

% Specify upper bounds
problem.ub01 = [3 2]';
problem.ub02 = [3 2]';
% ...

Tip

One could use problem.(sprintf('lb%02u',i)) in an i-indexed loop to set the parametric bounds more easily in the MATLAB client. Similarly, the parametric bounds for the stages can be set using problem["{:02d}".format(i+1)] in Python. Alternatively, consider using the option stack_parambounds, described below.

If the model.lbidx and model.ubidx fields vary for different stages use cell arrays of length \(N\). From Release 3.0.1, the parametric bounds can be stacked on one same array covering all stages. To enable this behavior, users need to set the following code-generation option:

codeoptions.nlp.stack_parambounds = 1;

This option is effective for both the PDIP_NLP and SQP_NLP solve methods and works with bounds on variables and inequalities. At run-time, users can then write

% Lower and upper bounds stacked over all stages
problem.lb = [0 0 0 0 ...];
problem.ub = [3 2 3 2 ...];

Alternatively, if you want to use the same bounds across all stages:

problem.lb = repmat([0, 0], 1, model.N);
problem.ub = repmat([3, 2], 1, model.N);

9.2.7. Variations

9.2.7.1. Varying dimensions, parameters, constraints, or functions

The example described above has the same dimensions, bounds and functions for the whole horizon. One can define varying dimensions using arrays and varying bounds and functions using MATLAB cell arrays or Python lists. For instance, to remove the first and second variables from the last stage one could write the following:

for i = 1:model.N-1
  model.nvar(i) = 6;
  model.objective{i} = @(z) -100*z(4) + 0.1*z(1)^2 + 0.01*z(2)^2;
  model.lb{i} = [-5, -1, -3, 0, 0, 0];
  model.ub{i} = [+5 , +1, 0, 3, 2, +pi];
  if i < model.N-1
    model.E{i} = [zeros(4, 2), eye(4)];
  else
    model.E{i} = eye(4);
  end
end

model.nvar(model.N) = 4;
model.objective{model.N} = @(z) -100*z(2);
model.lb{model.N} = [-3, 0, 0, 0];
model.ub{model.N} = [ 0, 3, 2, +pi];

It is also typical for model predictive control problems (MPC) that only the last stage differs from the others (excluding the initial condition, which is handled separately). Instead of defining cell arrays as above for all stages, FORCESPRO offers the following shorthand notations that alter the last stage:

  • nvarN: number of variables in last stage

  • nparN: number of parameters in last stage

  • objectiveN: objective function for last stage

  • EN: selection matrix \(E\) for last stage update

  • nhN: number of inequalities in last stage

  • ineqN: inequalities for last stage

Add any of these fields to the model struct/object to override the default values, which is to make everything the same along the horizon. For example, to add a terminal cost that is a factor 10 higher than the stage cost:

model.objectiveN = @(z) 10*model.objective(z);

9.2.7.2. Providing analytic derivatives

The algorithms inside FORCESPRO need the derivatives of the functions describing the objective, equality and inequality constraints. The code generation engine uses algorithmic differentiation (AD) to compute these quantities. Instead, when analytic derivatives are available, the user can provide them using the fields model.dobjective, model.deq, and model.dineq.

Note that the user must be particularly careful to make sure that the provided functions and derivatives are consistent, for example:

model.objective = @(z) z(3)^2;
model.dobjective = @(z) 2*z(3);

The code generation system will not check the correctness of the provided derivatives.

9.2.8. Single precision callbacks

Evaluating objective function, dynamics and constraints as well as their respective derivatives may take a significant part of the overall solution time (both total and per iteration). In such situations solution time and memory consumption may be sped up by evaluating those functions in single, rather than double precision arithmetic. This can be done by specifying

codeoptions.callback_floattype = 'float';

Note that this will allow to run the NLP solver in mixed-precision arithmetic, where the callbacks are evaluated in single precision, but the overall algorithm in double precision. In order for this to work well, all callbacks functions need to be numerically well-posed and overall accuracy requirements of the solution must not be too high. In particular, when using that feature, you may need to reduce some of the accuracy settings (such as codeoptions.nlp.TolStat) by one or two orders of magnitude, see Accuracy requirements.

Note

Single precision callbacks are currently supported for legacy and chainrule integrators, but not yet for VDE integrators. Also, this features is currently only available via the MATLAB client.

9.3. Generating a solver

In addition to the definition of the NLP, solver generation requires an (optional) set of options for customization (see the Solver Options section for more information). Using the default solver options we generate a solver using:

% Get the default solver options
codeoptions = getOptions('FORCESNLPsolver');

% Generate solver
FORCES_NLP(model, codeoptions);

As the solver is generated, several files are downloaded into the current working directory of the calling script, including the compiled solver itself and MATLAB/Python interfaces for calling it.

Note

In the Python client, generate_solver() returns a solver object. This object can be used to call the solver. To get a solver object for a previously generated solver in some directory /path/to/solver, use:

import forcespro.nlp
solver = forcespro.nlp.Solver.from_directory('/path/to/solver')

9.3.1. Declaring outputs

By default, the solver returns the solution vector for all stages as multiple outputs. Alternatively, the user can pass a third argument to the function FORCES_NLP with an array that specifies what the solver should output. For instance, to define an output, named u0, to be the first two elements of the solution vector at stage 1, use the following commands:

output1 = newOutput('u0', 1, 1:2);
FORCES_NLP(model, codeoptions, output1);

Additionally, you can request that the solver returns the solution vectors for all different stages “stacked” together into a single vector, say called sol, by using the following commands:

output1 = newOutput('sol');
FORCES_NLP(model, codeoptions, output1);

Important

When using the MINLP solver and defining outputs, all integer variables need to be specified as custom outputs.

The dual variables at the solution returned by FORCESPRO provide useful information on the problem sensitivity. They can be exported from the nonlinear solver as well by giving the maps2const field one of the following values:

  • ‘nl_eq_dual’ for the dual variables associated to equality constraints

  • ‘nl_lb_var_dual’ for the dual variables associated to lower bounds on variables

  • ‘nl_ub_var_dual’ for the dual variables associated to upper bounds on variables

  • ‘nl_ip_ineq_dual’ for the dual variables associated to nonlinear inequalities

  • ‘nl_ineq_slack’ for the dual variables associated to slacks on nonlinear inequalities.

An example of exporting the marginals associated to nonlinear equalities is shown in the code snippet below.

outputs(4) = newOutput('dual_eq0', 1:model.N, 1:2, 'nl_eq_dual');

9.4. Calling the solver

After code generation has been successful, one can obtain information about the real-time data needed to call the generated solver by typing:

help FORCESNLPsolver

In Python, a previously generated solver can be loaded as follows:

import forcespro.nlp
solver = forcespro.nlp.Solver.from_directory("/path/to/generated/solver/")
solver.help()

9.4.1. Initial guess

The FORCESPRO NLP solver solves NLPs to local optimality, hence the resulting optimal solution depends on the initialization of the solver. One can also choose another initialization point when a better guess is available. The following code sets the initial point to be in the middle of all bounds:

x0i = model.lb +(model.ub - model.lb)/2;
x0 = repmat(x0i', model.N, 1);
problem.x0 = x0;

9.4.2. Initial and final conditions

If there are initial and/or final conditions on the optimization variables, the solver will expect the corresponding runtime fields:

problem.xinit = model.xinit;
problem.xfinal = model.xfinal;

Note that the Python client does not allow setting model.xinit or model.xfinal properties, as those are run-time parameters not needed at solver generation time.

9.4.3. Real-time parameters

Whenever there are any runtime parameters defined in the problem, i.e. the field npar is not zero, the solver will expect the following field containing the parameters for all the \(N\) stages stacked in a single vector:

problem.all_parameters = repmat(1.0, model.N, 1);

9.4.4. Tolerances as real-time parameters

From FORCESPRO 2.0 onwards, the NLP solver tolerances can be made real-time parameters, meaning that they do not need to be set when generating the solver but can be changed at run-time when calling the generated solver. The code-snippet below shows how to make the tolerances on the gradient of the Lagrangian, the equalities, the inequalities and the complementarity condition parametric. Essentially, when the tolerances are declared nonpositive at code-generation, the corresponding run-time parameter is created in the solver.

codeoptions.nlp.TolStat = -1; % Tolerance on gradient of Lagrangian
codeoptions.nlp.TolEq = -1;   % Tolerance on equality constraints
codeoptions.nlp.TolIneq = -1; % Tolerance on inequality constraints
codeoptions.nlp.TolComp = -1; % Tolerance on complementarity

Once the tolerance has been declared nonpositive and the solver has been generated, the corresponding parameter can be set at run-time as follows:

problem.ToleranceStationarity = 1e-1;
problem.ToleranceEqualities = 1e-1;
problem.ToleranceInequalities = 1e-1;
problem.ToleranceComplementarity = 1e-1;

Tip

We do not recommend changing the tolerance on the complementarity condition since it is used internally to update the barrier parameter. Hence loosening it may hamper the solver convergence.

9.4.5. Exitflags and quality of the result

Once all parameters have been populated, the MEX interface of the solver can be used to invoke it:

[output, exitflag, info] = FORCESNLPsolver(problem);

The possible exitflags are documented in Exitflag values. The exitflag should always be checked before continuing with program execution to avoid using spurious solutions later in the code. Check whether the solver has exited without an error before using the solution. For example, in MATLAB, we suggest to use an assert statement:

assert(exitflag == 1, 'Some issue with FORCESPRO solver');

Besides the exitflag, the solver also returns an information structure containing detailed KPIs on the solver performance. All its entries are listed and explained in Table 9.1. Those values may also be useful in case the solver exitflag has been 0, which means the solver did not fail but reached the maximum number of iterations. In that case, one should at least check whether the “solution” returned is sufficiently feasible. This can be done by examining res_eq and res_ineq, respectively, to check whether equality and inequality constraints are satisfied up to a sufficiently small tolerance. The exact tolerances to check may be strongly application dependent.

Note

Applying a premature “solution” returned along with an exitflag of 0 to control your system may have undesired effects to the behavior of that system. Only do so if you fully understand what you are doing.

Table 9.1 Info struct entries

Fieldname

Description

it

Number of solver iterations that led to this result

res_eq

Maximum norm of equality constraint residual

res_ineq

Maximum norm of inequality constraint residual

rsnorm

Maximum norm of stationarity condition

rcompnorm

Maximum norm of violations of the complementarity conditions

pobj

Primal objective value (as provided by the user)

mu

Duality measure

solvetime

Time needed to run the solver (wall clock time)

fevalstime

Time needed just for function evaluations of all user callbacks inside the solver (wall clock time)

solver_id

Solver ID of generated solver

9.5. External function evaluations in C

This approach allows the user to integrate existing efficient C implementations to evaluate the required functions and their derivatives with respect to the stage variable. This gives the user full flexibility in defining the optimization problem. In this case, the functions do not necessarily have to be differentiable, although the convergence of the algorithm is not guaranteed if they are not. When following this route the user does not have to provide MATLAB code to evaluate the objective or constraint functions. However, the user is responsible for making sure that the provided derivatives and function evaluations are coherent. The FORCESPRO NLP code generator will not check this.

9.5.1. Interface

9.5.1.1. Expected function signature

There are two ways in which a user can supply custom functions written in C:

  • Either she can supply all functions (model.objective, model.eq, model.ineq etc.)

  • Or she can supply one or a few of these together with its differential/Jacobian.

Depending on the case, the user will have to supply different information when generating a FORCESPRO solver.

When supplying all functions, the user will have supply a single C function having the following signature:

int myfunctions (
    double *x, /* primal vars */
    double *y, /* eq. constraint multipliers */
    double *l, /* ineq . constraint multipliers */
    double *p, /* runtime parameters */
    double *f, /* objective function ( incremented in this function ) */
    double *nabla_f , /* gradient of objective function */
    double *c, /* dynamics */
    double *nabla_c , /* Jacobian of the dynamics ( column major ) */
    double *h, /* inequality constraints */
    double *nabla_h , /* Jacobian of inequality constraints ( column major ) */
    double *H, /* Hessian ( column major ) */
    int stage, /* stage number (0 indexed ) */
    int iteration, /* solver iteration count */
    int threadID /* thread id */
)

If instead the user wants to supply just a single function she will have to supply a single C function having the following signature:

void function (
    const double * const x, /* primal vars */
    const double * const p, /* runtime parameters */
    double * const zeroOrderFcn, /* Zero order function */
    double * const firstOrderFcn , /* first order Fcn */
    const int stage, /* stage number (0 indexed ) */
    const int threadID /* thread number */
)

where zeroOrderFcn and firstOrderFcn denote the function which the user wants to supply, together with its differential/Jacobian respectively. E.g. if the user were to add the objective function and its differential externally, the function might look as follows:

void objective (
    const double * const x, /* primal vars */
    const double * const p, /* runtime parameters */
    double * const obj, /* objective function */
    double * const nabla_obj, /* jacobian of objective fcn */
    const int stage, /* stage number (0 indexed ) */
    const int threadID /* thread number */
)

Note

External C-functions should have the same name as the file it is contained in, minus the file extension. E.g. in the above example the source file containing the definition of the function objective would have to have the name objective.c. If all functions are provided as external C functions through the FORCESPRO Python client, then one can provide a different name for the function and the file.

9.5.1.2. Custom data structures as parameters

If you have an advanced data structure that holds the user-defined run-time parameters, and you do not want to serialize it into an array of doubles to use the interface above, you can invoke the option:

codeoptions.customParams = 1;

options.customParams = 1

When doing this, it is important to note that run-time parameters can only be passed to externally provided functions. In particular, if some but not all function evaluations are provided externally, one will have to set model.npar = 0. When using custom parameters, if all functions are provided externally, the expected function signature is:

int myfunctions (
    double *x, /* primal vars */
    double *y, /* eq. constraint multipliers */
    double *l, /* ineq . constraint multipliers */
    void *p, /* runtime parameters (passed as void pointer) */
    double *f, /* objective function ( incremented in this function ) */
    double *nabla_f , /* gradient of objective function */
    double *c, /* dynamics */
    double *nabla_c , /* Jacobian of the dynamics ( column major ) */
    double *h, /* inequality constraints */
    double *nabla_h , /* Jacobian of inequality constraints ( column major ) */
    double *H, /* Hessian ( column major ) */
    int stage, /* stage number (0 indexed ) */
    int iteration, /* Solver iteration count */
    int threadID /* Thread id */
)

If instead only some of the functions are provided, the expected function signature of these is

void function (
    const double * const x, /* primal vars */
    void * const p, /* runtime parameters passed as void pointer */
    double * const zeroOrderFcn, /* Zero order function */
    double * const firstOrderFcn , /* first order Fcn */
    const int stage, /* stage number (0 indexed ) */
    const int threadID /* thread number */
)

At run time you can then pass arbitrary data structures to your own function by setting the pointer in the params struct:

myData p;   /* define your own parameter structure */
/* ... */   /* fill it with data */

/* Set parameter pointer to your data structure */
mySolver_params params; /* Define solver parameters */
params.customParams = &p;

/* Call solver (assuming everything else is defined) */
mySolver_solve(&params, &output, &info, stdout, &external_func);

Note

Setting customParams to 1 will disable building high-level interfaces. Only C header- and source files will be generated. As a consequence, if CasADi is used to generate some of the function evaluations, the generated source code will have to be compiled by the user.

Note

When using custom parameters, generating callback code automatically is only supported for CasADi 3.5.x only.

9.5.2. Supplying function evaluation information

In MATLAB, if the user wants to supply all functions externally she can let the code generator know about the path to the C files implementing the necessary function as follows:

model.extfuncs = 'C/myfunctions.c';

Alternatively she could supply either one of the functions model.objective, model.eq or model.ineq together with its differential by specifying the path to the corresponding C source file in the corresponding field of model.extfuncs as follows:

model.extfuncs.objective = 'C/myobjective.c'; % adding model.objective externally %
model.extfuncs.dynamics = 'C/mydynamics.c'; % adding model.eq externally %
model.extfuncs.inequalities = 'C/myinequalities.c'; % adding model.ineq externally %

As noted above, FORCESPRO derives the function name used for the callback from the file name; the function must therefore have the same name as the file in which it is contained.

In Python, if the user wishes to add ALL functions externally she would use a ExternalFunctionModel as follows:

model = forcespro.nlp.ExternalFunctionModel(50)
model.add_auxiliary(["helper_functions.c", "compiled_helper_functions.obj"])
model.set_main_callback("myfunctions.c", function="myfunctions")

Herein, the add_auxiliary() method is used to add any helper C source files or object files that should be compiled and linked against, and the set_main_callback() function is used to define the path to a C source file or compiled object file, as well as the name of an exported function that conforms to the call signature given above. This function will be used to evaluate any nonlinear constraints and the objective function.

Alternatively, in order to add a single function externally the user would use the SymbolicModel and add the C source files containing code for the external functions through model.extfuncs as follows:

model = forcespro.nlp.SymbolicModel(50)
model.extfuncs.objective = "myobjective.c"
model.extfuncs.dynamics = "mydynamics.c"
model.extfuncs.inequalities = "myinequalities.c"
model.add_c_source("extra_source.c")

where the extra_source.c are additional C source files needed for evaluating some or all of the external functions.

9.5.3. Rules for function evaluation code

The contents of the function have to follow certain rules. We will use the following example to illustrate them:

/* cost */
if (f)
{ /* notice the increment of f */
    (*f) += -100*x[3] + 0.1* x[0]*x[0] + 0.01*x [1]*x [1];
}
/* gradient - only nonzero elements have to be filled in */
if ( nabla_f )
{
    nabla_f [0] = 0.2*x[0];
    nabla_f [1] = 0.02*x[1];
    nabla_f [3] = -100;
}

/* eq constr */
if (c)
{
    vehicle_dynamics (x, c);
}
/* jacobian equalities ( column major ) */
if ( nabla_c )
{
    vehicle_dynamics_jacobian (x, nabla_c );
}

/* ineq constr */
if (h)
{
    h[0] = x [2]*x[2] + x[3]*x [3];
    h[1] = (x[2]+2)*(x[2]+2) + (x[3] -2.5)*(x[3] -2.5);
}
/* jacobian inequalities ( column major )
- only non - zero elements to be filled in */
if ( nabla_h )
{
    /* column 3 */
    nabla_h [4] = 2*x [2];
    nabla_h [5] = 2*x[2] + 4;
    /* column 4 */
    nabla_h [6] = 2*x [3];
    nabla_h [7] = 2*x[3] - 5;
}

Notice that every function evaluation is only carried out if the corresponding pointer is not null. This is used by the FORCESPRO NLP solver to call the same interface with different pointers depending on the functions that it requires.

9.5.4. Matrix format

Matrices are assumed to be stored in dense column major format. However, only the non-zero components need to be populated, as FORCESPRO NLP makes sure that the arrays are initialized to zero before calling this interface.

9.5.5. Multiple source files

The use of multiple C files is also supported. In the example above, the functions dynamics and dynamics_jacobian are defined in another file and included as external functions using:

extern void dynamics ( double *x, double *c);
extern void dynamics_jacobian ( double *x, double *J);

In MATLAB, to let the code generator know about the location of these other files use a string with spaces separating the different files. In Python, use the add_auxiliary() method:

codeoptions.nlp.other_srcs = 'C/dynamics.c';

9.5.6. Stage-dependent functions

Whenever the cost function in one of the stages is different from the standard cost function \(f\), one can make use of the argument stage to evaluate different functions depending on the stage number. The same applies to all other quantities.

9.5.7. External function return values

By default, FORCESPRO does not check the return value of the main callback function. If you write your external function so that it returns:

  • a non-negative value in case of success,

  • a negative value between -99 and -1 in case of failure,

you can set the FORCESPRO code option codeoptions.callback_check_status = 1 to make the solver check these return values and stop should a value be negative. In the latter case, the solver returns ret - 200, given ret as the external function return value. Refer to section Exitflag values for a full description of the relevant exitflags. If the callbacks are evaluated concurrently (which is the case if codeoptions.parallel > 0), the solver exitflag refers to the most negative return value from external function evaluations.

Note

You need to verify yourself the correctness of the external function signature (as documented here) as this can’t be verified automatically.

9.6. Calling the nonlinear functions from MATLAB or Python

From FORCESPRO version 5.0.1 onwards it is possible to evaluate the nonlinear functions and their derivatives directly in MATLAB or Python. This can be very useful for debugging a FORCESPRO solver. This in particular can be useful when calling the generated solver returns an exitflag \(-6\), \(-8\) or \(-10\). For help on how to do this, see Real-time SQP Solver: Robotic Arm Manipulator (MATLAB & Python) and Issues when running a generated solver.

9.6.1. Calling the nonlinear functions from MATLAB

When generating a FORCESPRO solver one can request the FORCESPRO client to also generate a MEX entry point to any of the nonlinear functions (equality constraints, inequality constraints and objective function) so that one can evaluate these along with their derivatives directly in MATLAB. This is done via the following three options:

  • dynamics: In order to generate a MEX interface for the equality constraints, set codeoptions.MEXinterface.dynamics = 1.

  • inequalities: In order to generate a MEX interface for the inequality constraints, set codeoptions.MEXinterface.inequalities = 1.

  • objective: In order to generate a MEX interface for the objective function, set codeoptions.MEXinterface.objective = 1.

The generated MEX entry point carries the name <solvername>_dynamics, <solvername>_inequalities or <solvername>_objective depending on the case.

The first argument to each of these generated MEX entry points is the primal variable \(z_k\) and the second argument is the real-time parameter \(p_k\) (see Canonical problem for discrete-time dynamics), both passed as column vectors. Additionally a third argument can be passed, which indicates the stage (1-based) at which the function should be evaluated. Each of the generated MEX entry points output two variables: The first is the constraint/objective evaluated at the supplied point and the second is its jacobian with respect to \(z_k\). E.g. in order to generate a solver named FORCESsolver and generate a MEX entry point for the inequality constraints, one would proceed as follows:

model = getModel();
codeoptions = getOptions('FORCESsolver');
codeoptions.MEXinterface.inequalities = 1;
FORCES_NLP(model, codeoptions);

Now one can evaluate the inequality constraints on stage \(8\) at the point \(z_8 = (4.0, 3.0, 2.4)^T\) with real-time parameters given by \(p_8 = (3.4, 2.1)^T\) by doing

z = [4.0; 3.0; 2.4];
p = [3.4; 2.1];
ineq, jacineq = FORCESsolver_inequalities(z,p,8);

which stores the values of the inequality constraints in ineq and their jacobian in jacineq.

Note

If there are no real-time parameters one simply passes an empty array: ineq, jacineq = FORCESsolver_inequalities(z,[],8)

9.6.2. Calling the nonlinear functions from Python

When generating a solver one obtains a solver object

solver = model.generate_solver(options)

or

import forcespro.nlp
solver = forcespro.nlp.Solver.from_directory('/path/to/solver')

and one can directly evaluate any of the nonlinear functions (equality constraints, inequalities or objective function) via the solver object’s methods. The methods for doing so are dynamics, ineq and objective. Each of these functions take as a first input the primal variable \(z_k\) and the real-time parameter \(p_k\) as a second input, if the problem formulation has real-time parameters (see Canonical problem for discrete-time dynamics). All of these inputs are expected to be passed as numpy arrays. Recall that the way in which states and inputs are “organized” into \(z_k\) is determined by model.E. Additionally a stage input can be given in order to compute the dynamics at a given stage (it is not necessary to supply this argument if the same dynamics are used throughout the entire horizon). This stage input should give the 0-based number of the corresponding stage. The first output of each of these function is the evaluated constraint/objective and the second output is the derivative with respect to \(z_k\). E.g. in order to evaluate the equality constraints at stage number \(7\) on the primal variable \(z_8 = (4.0, 3.0, 2.4)^T\) in a formulation where there are no real-time parameters, one would do as follows:

import numpy as np
import forcespro.nlp

solver = forcespro.nlp.Solver.from_directory('/path/to/solver')
z = np.array([4.0, 3.0, 2.4])
c, jacc = solver.dynamics(z, stage=7)

This stores the value of the equality constraints in the numpy array c and its jacobian in the numpy array jacc.

9.7. Mixed-integer nonlinear solver

From FORCESPRO 1.8.0, mixed-integer nonlinear programs (MINLPs) are supported. This broad class of problems encompasses all nonlinear programs with some integer decision variables.

This interface is provided with Variant L of FORCESPRO.

9.7.1. Writing a mixed-integer model

In order to use this feature, the user has to declare lower and upper bounds on all variables as parametric, as shown in the code below.

model.lb = [];
model.ub = [];

The user is then expected to provide lower and upper bounds as run-time parameters. FORCESPRO switches to the MINLP solver as soon as some variables are declared as integers in any stage. This information can be provided to FORCESPRO via the intidx array at every stage. An example is shown below.

%% Add integer variables to existing nonlinear model
for s = 1:5
    model.intidx{s} = [4, 5, 6];
end

In the above code snippet, the user declares variables 4, 5 and 6 (3, 4 and 5 in Python’s zero-based indexing) as integers from stage 1 to 5 (stages 0 to 4 in Python’s zero-based indexing). The values that can be taken by an integer variable are derived from its lower and upper bounds. For instance, if the variable lies between -1 and 1, then it can take integer values -1, 0 or 1. If a variable has been declared as integer and does not have lower or upper bounds, FORCESPRO raises an exception during code generation. Stating that a variable has lower and upper bounds should be done via the arrays lbidx and ubidx. For instance, in the code below, variables 1 to 6 (0 to 5 in Python) in stage 1 (0) have lower and upper bounds, which are expected to be provided at run-time.

model.lbidx{1} = 1 : 6;
model.ubidx{1} = 1 : 6;

The FORCESPRO MINLP algorithm is based on the well-known branch-and-bound algorithm but comes with several customization features which generally help for improving performance on some models by enabling the user to provide application specific knowledge into the search process. At every node of the search tree, the FORCESPRO nonlinear solver is called in order to compute a solution of a relaxed problem. The generated MINLP solver code can be customized via the options described in Table 9.2, which can be changed before running the code generation.

One of the salient features of the MINLP solver is that the branch-and-bound search can be run in parallel on several threads. Therefore the search is split in two phases. It starts with a sequential branch-and-bound and switches to a parallelizable process when the number of nodes in the queue is sufficiently high. The node selection strategy can be customized in both phases, as described in Table 9.2.

Table 9.2 FORCESPRO MINLP solver options

Code generation setting

Values

Default

minlp.int_gap_tol

Any value \(\geq 0\)

0.001

minlp.max_num_nodes

Any value \(\geq 0\)

10000

minlp.seq_search_strat

'BEST_FIRST', 'BREADTH_FIRST 'DEPTH_FIRST'

'BEST_FIRST'

minlp.par_search_strat

'BEST_FIRST', 'BREADTH_FIRST', 'DEPTH_FIRST'

'BEST_FIRST'

minlp.max_num_threads

Any nonnegative value preferably smaller than 8

4

minlp.output_relaxation

\(0\) or \(1\)

\(0\)

  • The minlp.int_gap_tol setting corresponds to the final optimality tolerance below which the solver is claimed to have converged. It is the difference between the objective incumbent, which is the best integer feasible solution found so far and the lowest lower bound. As the node problems are generally not convex, it can be expected to become negative. FORCESPRO claims convergence to a local minimum only when the integrality gap is nonnegative and below the tolerance minlp.int_gap_tol.

  • The minlp.max_num_nodes setting is the maximum number of nodes which can be explored during the search.

  • The minlp.seq_search_strat setting is the search strategy which is used to select candidate nodes during the sequential search phase.

  • The minlp.par_search_strat setting is the search strategy which is used to select candidate nodes during the parallelizable search phase.

  • The minlp.max_num_threads setting is the maximum number of threads allowed for a parallel search. The actual number of threads on which the branch-and-bound algorithm can be run can be set as a run-time parameter, as described below.

  • The minlp.output_relaxation setting enables users to export the primal outputs of the root relaxation. With this option set to 1, the server automatically generates one additional output for every defined output. The name of the root relaxation output is the name of the output followed by _relax.

Note

MINLP formulations are currently not supported on MacOS. Moreover, they cannot be used in combination with an SQP algorithm (see Sequential quadratic programming algorithm) nor with CasADi MX expressions (see Automatic differentiation expression class).

The FORCESPRO MINLP solver also features settings which can be set at run-time. These are the following:

  • minlp.numThreadsBnB, the number of threads used to parallelize the search. Its default value is 1, if not provided by the user.

  • minlp.solver_timeout, the maximum amount of time allowed for completing the search. Its default value is 10.0 seconds, if not set by the user.

  • minlp.parallelStrategy, the method used for parallelizing the mixed-integer search (from FORCESPRO 1.9.0). Value 0 (default) corresponds to a single priority queue shared between threads. Value 1 corresponds to having each thread managing its own priority queue.

9.7.2. Mixed-integer solver customization via user callbacks

For advanced users, the mixed-integer branch-and-bound search can be customized after the rounding and the branching phases. In the rounding phase, an integer feasible solution is computed after each relaxed problem solve. The user is allowed to modify the rounded solution according to some modelling requirements and constraints. This can be accomplished via the postRoundCallback_template.c file provided in the FORCESPRO client. This callback is applied at every stage in a loop and updates the relaxed solution stage-wise. It needs to be provided before code generation, as shown in the following code snippet.

%% Add post-rounding callback to existing model
postRndCall = fileread('postRoundCallback_template.c'); % The file name can be changed by the user
model.minlpPostRounding = postRndCall;

The branching process can be customized in order to discard some nodes during the search. To do so, the user is expected to overwrite the file postBranchCallback_template.c and pass it to FORCESPRO before generating the MINLP solver code.

%% Add as post-branching callbacks as you want
postBranchCall_1 = fileread('postBranchCallback_template_1.c');
postBranchCall_2 = fileread('postBranchCallback_template_2.c');
postBranchCall_3 = fileread('postBranchCallback_template_3.c');
model.minlpPostBranching{1} = postBranchCall_1;
model.minlpPostBranching{2} = postBranchCall_2;
model.minlpPostBranching{3} = postBranchCall_3;

In each of those callbacks, the user is expected to update the lower and upper bounds of the sons computed during branching given the index of the stage in which the branched variables lies, the index of this variable inside the stage and the relaxed solution at the parent node.

9.7.3. Providing a guess for the incumbent

Internally, the mixed-integer branch-and-bound computes an integer feasible solution by rounding. Moreover, since version 1.9.0, users are allowed to provide an initial guess for the incumbent. At code-generation, the following options need to be set:

  • minlp.int_guess, which tells whether an integer feasible guess is provided by the user (value 1). Its default value is 0.

  • minlp.int_guess_stage_vars, which specifies the indices of the integer variables that are user-initialized within one stage (MATLAB based indexing). If minlp.int_guess = 1, a parameter int_guess needs to be set at every stage. An example can be found there Mixed-integer nonlinear solver: F8 Crusader aircraft.

Another important related option is minlp.round_root. If set to 1, the solution of the root relaxation is rounded and set as incumbent if feasible. Its default value is 1. The mixed-integer solver behavior differs depending on the combinations of options. The different behaviors are listed below.

  • If minlp.int_guess = 0 and minlp.round_root = 1, then the solution of the root relaxation is taken as incumbent (if feasible). This is the default behavior.

  • If minlp.int_guess = 1 and minlp.round_root = 0, then the incumbent guess provided by the user is tested after the root solve. If feasible, it is taken as incumbent. Note that the user is allowed to provide guesses for a few integers per stage only. In this case, the other integer variables are rounded to the closest integer.

  • If minlp.int_guess = 1 and minlp.round_root = 1, then the rounded solution of the root relaxation and the user guess are compared. The best integer feasible solution in terms of primal objective is then taken as incumbent.

This feature is illustrated in Example Mixed-integer nonlinear solver: F8 Crusader aircraft. The ability of providing an integer guess for the incumbent is a key feature to run the mixed-integer solver in a receding horizon setting.

9.8. Sequential quadratic programming algorithm

The FORCESPRO real-time sequential quadratic programming (SQP) algorithm allows one to solve problems of the type specified in the section High-level Interface. The algorithm iteratively solves a convex quadratic approximations of the (generally non-convex) problem. Moreover, the solution is stored internally in the solver and used as an initial guess for the next time the solver is called. This and other features enables the solver to have fast solvetimes (compared to the interior point method), particularly suitable for MPC applications where the sampling time or the computational power of the hardware is small.

Important

The SQP algorithm currently only supports affine inequalities. This means that all the inequality functions \(h_k, k=1,...,N\) from (9.1) must be affine functions of the variable \(z_k\) (not necessarily of \(p_k\)).

Moreover, the SQP algorithm currently does not support problems comprising final equality constraints (specified via model.xfinalidx).

9.8.1. How to generate a SQP solver

To generate a FORCESPRO sequential quadratic programming real-time iteration solver one sets

codeoptions.solvemethod = 'SQP_NLP';

(see Generating a solver). In addition to the general code options specified in the previous section here are some of the important code options one can use to customize the generated SQP solver.

By default the FORCESPRO SQP solver solves a single convex quadratic approximation. This accomplishes a fast solvetime compared to a “full” sequential quadratic programming solver (which solves quadratic approximations to the nonlinear program until a KKT point is reached). The user might prefer to manually allow the SQP solver to solve multiple quadratic approximations: By setting

codeoptions.sqp_nlp.maxqps = k;

for a positive integer k one allows the solver to solve k quadratic approximations at every call to the solver. In general, the more quadratic approximations which are solved, the higher the control performance. The tradeoff is that the solvetime also increases.

9.8.2. Different SQP variants

In addition to the default FORCESPRO SQP solver (also referred to as SQP General here), FORCESPRO supports a new SQP Fast solver from version 6.0.0. An advantage of the fast SQP solver is that it is typically faster and can be made to run on more low-level hardware. An advantage of the general (default) SQP solver is that it is very robust. See Tuning the SQP Fast solver for a detailed discussion on how to generate a SQP Fast solver and see High-level interface: Path tracking example (MATLAB) for an example of how to use the SQP Fast solver.

Note

The SQP Fast solver is only supported when using a Gauss-Newton hessian approximation and disabling the line search (see The hessian approximation and line search settings). The SQP Fast solver is currently only supported via the MATLAB client of FORCESPRO.

9.8.3. Tuning the SQP Fast solver

One of the novel features of the fast SQP solver is that it can be tuned/tailored to achieve optimal performance on a specific application. FORCESPRO provides a tool for “autotuning” the solver (see Autotuner). The main step in using this tool is to collect training problems which can be used to tune the solver. The standard way to do this is to first generate an SQP General (codeoptions.sqp_nlp.qp_method = "general", which is also the default value) solver, perform a closed-loop simulation with this solver while caching/saving all problem instances (i.e. the inputs to the solver at run-time). In a second step a SQP Fast solver is generated and automatically tuned by calling ForcesGenerateSqpFastSolver. The complete workflow is as follows:

tuningoptions = ForcesAutotuneOptions(problems)
ForcesGenerateSqpFastSolver(model, codeoptions, tuningoptions);

where problems is a cell array of problems collected from the simulation with the SQP General solver. The autotuning procedure allows for quite a bit of customization which can be specified through the ForcesAutotuneOptions object (see Autotuner Options), and one can also pass multiple problems for the tuning process which should result in better tuned parameters. See High-level interface: Path tracking example (MATLAB) for an example of how to tune the SQP Fast solver.

9.8.4. The hessian approximation and line search settings

The SQP code generation currently supports two different types of hessian approximations. A good choice of hessian approximation can often improve the number of iterations required by the solver and thereby its solvetime. The default option for a SQP solver is the BFGS hessian approximation. When the objective function of the optimization problem is a least squares cost it is often beneficial to use the Gauss-Newton hessian approximation instead. To enable this option one proceeds as specified in the sections Hessian approximation and Gauss-Newton options. When the Gauss-Newton hessian approximation is chosen one can also disable the the internal linesearch by setting

codeoptions.sqp_nlp.use_line_search = 0;

A linesearch is required to ensure global convergence of an SQP method, but is not needed in a real-time context when a Gauss-Newton hessian approximation is used.

Note

One cannot disable the line search when using the BFGS hessian approximation.

9.8.5. Controlling the initial guess at run-time

Upon the first call to the generated FORCESPRO SQP solver one needs to specify a primal initial guess (problem.x0, see also Initial guess). The default behavior of the FORCESPRO SQP solver is to use the solution from the previous call as initial guess in every subsequent call to it. However, one can also manually set an initial guess in subsequent calls to the solver. Wether a manual initial guess (provided through problem.x0) will be used or the internally stored solution from the previous call will be used can be controlled by the field problem.reinitialize of the problem struct which is passed as an argument to the solver when it is called.

The reinitialize field can take two values: 0 or 1. For the default usage of the solver

problem.reinitialize = 0;

should be used. This choice results in the solver using the solution from the previous call as initial guess. This feature is useful when running the real-time iteration scheme because it ensures that the initial guess is close to the optimal solution. If you want to specify an initial guess at run-time, you will need to set

problem.reinitialize = 1;

So in summary: The first time the solver is called the initial guess the solver will use has to be provided by problem.x0. In all subsequent calls the solver will only make use of problem.x0 as its initial guess if problem.reinitialize = 1.

9.8.6. Additional code options specific to the SQP-RTI solver

In addition to the above codeoptions, the following options are specific to the SQP algorithm. Each of these options can be supplied when generating a solver as a field of codeoptions.sqp_nlp (e.g. codeoptions.sqp_nlp.TolStat).

Table 9.3 SQP specific codeoptions

option

Possible values

Default value

Description

TolStat

positive

\(10^{-6}\)

Set the stationarity tolerance required for terminating the algorithm (the tolerance required to claim convergence to a KKT point).

TolEq

positive

\(10^{-6}\)

Set the feasibility tolerance required for terminating the algorithm (the tolerance required to claim convergence to a feasible point).

reg_hessian

positive

\(5\cdot 10^{-9}\)

Set the level of regularization of the hessian approximation (often increasing this parameter can help if the SQP solver returns exitflag -8 for your problem)

qpinit

\(0\) or \(1\)

\(0\)

Set the initialization strategy for the internal QP solver. \(0\) = cold start and \(1\) = centered start. See also Solver Initialization (note however, that for the SQP solver qpinit=2 is not possible).

In addition to these options one can also specify the maximum number of iterations the internal QP solver is allowed to run in order to solve the quadratic approximation. If one wishes the QP solver use no more than k iterations to solve a problem one sets

codeoptions.maxit = k;

Note

The SQP fast solver currently does not support parallel execution, i.e. setting codeoptions.parallel will have no effect.

9.9. Differences between the MATLAB and the Python client

The Python NLP interface is largely similar to the MATLAB interface, but does come with some language- and implementation-specific differences.

  • All indices in the problem formulation are expected to be 0-based in Python, as is usual in this language. This does not include the indices of the generated solver, however, where outputs are named x01, x02, … as in MATLAB. Thus, the problem formulation before generation requires 0-based indices, whereas the returned solver from the server uses 1-based indices. This also does not apply to the low-level Python interface, where indices are 1-based even in the model formulation.

  • In the Python client, different model objects must be used when using external functions or symbolic expressions, namely nlp.ExternalFunctionModel() and nlp.SymbolicModel(). Furthermore, if the high-level interface is to be used for convex problems, this is only possible using the nlp.ConvexSymbolicModel(). This is different from the MATLAB client, where the FORCES_NLP function accepts problems of any kind and switches to the appropriate solver automatically.

  • When using the Python client with a nlp.SymbolicModel(), the C code generated for symbolic expressions is currently not entirely identical to the code generated by MATLAB. While the actual expression evaluation code generated by CasADi is the same, the structure of the files varies. Specifically, the MATLAB client creates individual C files for each problem stage with distinct symbolic expressions (leading to varying file names when changing the problem horizon) whereas all functions are gathered in one file in the Python client. Yet, the Python client does add one additional file for the FORCESPRO-CasADi glue code, which is not present when using the MATLAB client. Lastly, function names of the evaluation functions differ.

    If you want to get the same code for MATLAB and Python, you must generate the CasADi C code from one of both clients and then supply this code as an external function in the other client.

9.10. Examples