FunctionApproximation.Problem Class

Namespace: FunctionApproximation

Object defining the function to approximate, or the lookup table to optimize

Description

The FunctionApproximation.Problem object defines the math function, function handle, cfit object, or Simulink^® block to approximate with a lookup table, or the lookup table block to optimize. After defining the problem, use the solve method to generate a FunctionApproximation.LUTSolution object that contains the approximation.

Creation

problem = FunctionApproximation.Problem() creates a FunctionApproximation.Problem object with default property values. When no function input is provided, the FunctionToApproximate property is set to 'sin'.

problem = FunctionApproximation.Problem(function) creates a FunctionApproximation.Problem object to approximate the math function, function handle, cfit object, or Simulink block, or the lookup table block to optimize, specified by function.

Input Arguments

function — Function, object, or block to approximate, or lookup table block to optimize

'sin' (default) | math function | function handle | cfit object | Math Function block | Subsystem block | Lookup Table block

Function, object, or block to approximate, or the lookup table block to optimize, specified as a math function, function handle, cfit (Curve Fitting Toolbox) object, Simulink block or subsystem, or one of the lookup table blocks (for example, 1-D Lookup Table, n-D Lookup Table).

If you specify a math function, a function handle, cfit object, or a Simulink block, the solve method generates a lookup table approximation of the input function or block.

If you specify one of the lookup table blocks, the solve method generates an optimized lookup table.

The MATLAB^® math functions supported for approximation are:

• 1./x

• 10.^x

• 2.^x

• acos

• acosh

• asin

• asinh

• atan

• atan2

• atanh

• cos

• cosh

• exp

• log

• log10

• log2

• sin

• sinh

• sqrt

• tan

• tanh

• x.^2

Function handles must be on the MATLAB search path, or approximation fails.

Tip

The process of generating a lookup table approximation is faster for a function handle than for a subsystem. If a subsystem can be represented by a function handle, it is faster to approximate the function handle.

If you specify a cfit object, use the fittype (Curve Fitting Toolbox) function to specify a library model to approximate. For a list of library models, see List of Library Models for Curve and Surface Fitting (Curve Fitting Toolbox).

Data Types: char | string | function_handle

Properties

FunctionToApproximate — Function, object, or block to approximate, or lookup table block to optimize

'sin' (default) | math function | function handle | cfit object | Math Function block | Subsystem block | Lookup Table block

Function, object, or block to approximate, or the lookup table block to optimize, specified as a math function, function handle, cfit (Curve Fitting Toolbox) object, Simulink block or subsystem, or one of the lookup table blocks (for example, 1-D Lookup Table, n-D Lookup Table).

If you specify a math function, a function handle, cfit object, or a Simulink block, the solve method generates a lookup table approximation of the input function or block.

If you specify one of the lookup table blocks, the solve method generates an optimized lookup table.

The MATLAB math functions supported for approximation are:

• 1./x

• 10.^x

• 2.^x

• acos

• acosh

• asin

• asinh

• atan

• atan2

• atanh

• cos

• cosh

• exp

• log

• log10

• log2

• sin

• sinh

• sqrt

• tan

• tanh

• x.^2

Function handles must be on the MATLAB search path, or approximation fails.

Tip

The process of generating a lookup table approximation is faster for a function handle than for a subsystem. If a subsystem can be represented by a function handle, it is faster to approximate the function handle.

If you specify a cfit object, use the fittype (Curve Fitting Toolbox) function to specify a library model to approximate. For a list of library models, see List of Library Models for Curve and Surface Fitting (Curve Fitting Toolbox).

Data Types: char | string | function_handle

NumberOfInputs — Number of inputs to function approximation

1 | 2 | 3

Number of inputs to approximated function. This property is inferred from the FunctionToApproximate property, therefore it is not a writable property.

If you are generating a Direct Lookup Table, the function to approximate can have no more than two inputs.

Data Types: double

InputTypes — Desired data types of inputs to function approximation

numerictype object | vector of numerictype objects | Simulink.Numerictype object | vector of Simulink.Numerictype objects

Desired data types of the inputs to the approximated function, specified as a numerictype, Simulink.Numerictype, or a vector of numerictype or Simulink.Numerictype objects. The number of InputTypes specified must match the NumberOfInputs.

Example: problem.InputTypes = ["numerictype(1,16,13)", "numerictype(1,16,10)"];

InputLowerBounds — Lower limit of range of inputs to function to approximate

scalar | vector

Lower limit of range of inputs to function to approximate, specified as a scalar or vector. If you specify inf, the InputLowerBounds used during the approximation is derived from the InputTypes property. The dimensions of InputLowerBounds must match the NumberOfInputs.

Data Types: single | double | int8 | int16 | int32 | int64 | uint8 | uint16 | uint32 | uint64 | fi

InputUpperBounds — Upper limit of range of inputs to function to approximate

scalar | vector

Upper limit of range of inputs to function to approximate, specified as a scalar or vector. If you specify inf, the InputUpperBounds used during the approximation is derived from the InputTypes property. The dimensions of InputUpperBounds must match the NumberOfInputs.

Data Types: single | double | int8 | int16 | int32 | int64 | uint8 | uint16 | uint32 | uint64 | fi

OutputType — Desired data type of the function approximation output

numerictype object | Simulink.Numerictype object

Desired data type of the function approximation output, specified as a numerictype or Simulink.Numerictype object. For example, to specify that you want the output to be a signed fixed-point data type with 16-bit word length and best-precision fraction length, set the OutputType property to "numerictype(1,16)".

Example: problem.OutputType = "numerictype(1,16)";

Options — Additional options and constraints to use in approximation

FunctionApproximation.Options object

Additional options and constraints to use in approximation, specified as a FunctionApproximation.Options object.

Methods

Public Methods

solve

Solve for optimized solution to function approximation problem

Examples

Create Problem Object to Approximate a Math Function

Create a FunctionApproximation.Problem object specifying a math function to approximate.

problem = FunctionApproximation.Problem('log')

problem = 
  1x1 FunctionApproximation.Problem with properties:

    FunctionToApproximate: @(x)log(x)
           NumberOfInputs: 1
               InputTypes: "numerictype(1,16,10)"
         InputLowerBounds: 0.6250
         InputUpperBounds: 15.6250
               OutputType: "numerictype(1,16,13)"
                  Options: [1x1 FunctionApproximation.Options]

Create Problem Object to Approximate a Function Handle

Create a FunctionApproximation.Problem object specifying a function handle that you want to approximate.

problem = FunctionApproximation.Problem(@(x,y) sin(x)+cos(y))

problem = 
  1x1 FunctionApproximation.Problem with properties:

    FunctionToApproximate: @(x,y)sin(x)+cos(y)
           NumberOfInputs: 2
               InputTypes: ["numerictype('double')"    "numerictype('double')"]
         InputLowerBounds: [-Inf -Inf]
         InputUpperBounds: [Inf Inf]
               OutputType: "numerictype('double')"
                  Options: [1x1 FunctionApproximation.Options]

The FunctionApproximation.Problem object, problem, uses default property values.

Set the range of the function inputs to be between 0 and 2*pi.

problem.InputLowerBounds = [0,0];
problem.InputUpperBounds = [2*pi, 2*pi]

problem = 
  1x1 FunctionApproximation.Problem with properties:

    FunctionToApproximate: @(x,y)sin(x)+cos(y)
           NumberOfInputs: 2
               InputTypes: ["numerictype('double')"    "numerictype('double')"]
         InputLowerBounds: [0 0]
         InputUpperBounds: [6.2832 6.2832]
               OutputType: "numerictype('double')"
                  Options: [1x1 FunctionApproximation.Options]

Create Problem Object to Optimize a Lookup Table Block

Create a FunctionApproximation.Problem object to optimize an existing lookup table.

openExample('simulink_automotive/ModelingAFaultTolerantFuelControlSystemExample',...
    'supportingfile','sldemo_fuelsys');
problem = FunctionApproximation.Problem('sldemo_fuelsys/fuel_rate_control/airflow_calc/Pumping Constant')

problem = 

  1×1 FunctionApproximation.Problem with properties:

    FunctionToApproximate: 'sldemo_fuelsys/fuel_rate_control/airflow_calc/Pumping Constant'
           NumberOfInputs: 2
               InputTypes: ["numerictype('single')"    "numerictype('single')"]
         InputLowerBounds: [50 0.0500]
         InputUpperBounds: [1000 0.9500]
               OutputType: "numerictype('single')"
                  Options: [1×1 FunctionApproximation.Options]

The software infers the properties of the problem object from the model.

Create Problem Object to Approximate a Curve Fitting Object

Create a FunctionApproximation.Problem object specifying a cfit object to approximate.

ffun = fittype('exp1');
cfun = cfit(ffun,0.1,0.2);
problem = FunctionApproximation.Problem(cfun)

problem = 
  1x1 FunctionApproximation.Problem with properties:

    FunctionToApproximate: [1x1 cfit]
           NumberOfInputs: 1
               InputTypes: "numerictype('double')"
         InputLowerBounds: -Inf
         InputUpperBounds: Inf
               OutputType: "numerictype('double')"
                  Options: [1x1 FunctionApproximation.Options]

Search for Floating-Point Solution to Function Approximation Problem

Since R2023a

This example shows how to search for pure floating-point solutions to the function approximation problem.

Create a FunctionApproximation.Problem object specifying a function to approximate.

problem = FunctionApproximation.Problem("sin");

Specify the input and output types to be a floating-point data type.

problem.InputTypes = [numerictype('Single')];
problem.OutputType = [numerictype('Single')];

Use the FunctionApproximation.Options object to specify word lengths that can be used in the lookup table approximation. To search for floating-point solutions, specify word lengths corresponding to a single-precision or double-precision data type.

problem.Options.WordLengths = 32;

Use the solve method to generate an approximation of the function.

solve(problem)

Searching for fixed-point solutions.

|  ID |  Memory (bits) | Feasible | Table Size | Breakpoints WLs | TableData WL | BreakpointSpecification |             Error(Max,Current) | 
|   0 |            128 |        0 |          2 |              32 |           32 |             EvenSpacing |     7.812500e-03, 1.000000e+00 |
|   1 |           1568 |        1 |         47 |              32 |           32 |             EvenSpacing |     7.812500e-03, 2.331257e-03 |
|   2 |           1536 |        1 |         46 |              32 |           32 |             EvenSpacing |     7.812500e-03, 2.434479e-03 |
|   3 |           1216 |        1 |         36 |              32 |           32 |             EvenSpacing |     7.812500e-03, 4.021697e-03 |
|   4 |           1184 |        1 |         35 |              32 |           32 |             EvenSpacing |     7.812500e-03, 4.265845e-03 |
|   5 |            832 |        1 |         24 |              32 |           32 |             EvenSpacing |     7.812500e-03, 6.421237e-03 |
|   6 |            800 |        1 |         23 |              32 |           32 |             EvenSpacing |     7.812500e-03, 7.061585e-03 |
|   7 |            448 |        0 |         12 |              32 |           32 |             EvenSpacing |     7.812500e-03, 4.009663e-02 |
|   8 |            608 |        0 |         17 |              32 |           32 |             EvenSpacing |     7.812500e-03, 1.884634e-02 |
|   9 |            704 |        0 |         20 |              32 |           32 |             EvenSpacing |     7.812500e-03, 8.071933e-03 |
|  10 |            736 |        0 |         21 |              32 |           32 |             EvenSpacing |     7.812500e-03, 8.607101e-03 |
|  11 |            768 |        1 |         22 |              32 |           32 |             EvenSpacing |     7.812500e-03, 7.243455e-03 |
|  12 |            128 |        0 |          2 |              32 |           32 |         EvenPow2Spacing |     7.812500e-03, 1.315148e+00 |
|  13 |           1152 |        1 |         18 |              32 |           32 |          ExplicitValues |     7.812500e-03, 7.812380e-03 |
|  14 |           1024 |        0 |         16 |              32 |           32 |          ExplicitValues |     7.812500e-03, 1.202238e-02 |
|  15 |           1152 |        0 |         18 |              32 |           32 |          ExplicitValues |     7.812500e-03, 1.068657e-02 |
|  16 |           1280 |        1 |         20 |              32 |           32 |          ExplicitValues |     7.812500e-03, 7.278687e-03 |
Searching for floating-point solutions.

|  17 |           1536 |        1 |         46 |              32 |           32 |             EvenSpacing |     7.812500e-03, 2.434489e-03 |
|  18 |            128 |        0 |          2 |              32 |           32 |         EvenPow2Spacing |     7.812500e-03, 1.315148e+00 |
|  19 |           1152 |        1 |         18 |              32 |           32 |          ExplicitValues |     7.812500e-03, 7.812365e-03 |
|  20 |           1024 |        0 |         16 |              32 |           32 |          ExplicitValues |     7.812500e-03, 1.202232e-02 |

Best Solution
|  ID |  Memory (bits) | Feasible | Table Size | Breakpoints WLs | TableData WL | BreakpointSpecification |             Error(Max,Current) |
|  11 |            768 |        1 |         22 |              32 |           32 |             EvenSpacing |     7.812500e-03, 7.243455e-03 |

ans = 
  1x1 FunctionApproximation.LUTSolution with properties:

          ID: 11
    Feasible: "true"

The solve method returns all feasible solutions. In the table, fixed-point solutions are returned first, followed by floating-point solutions. The Lookup Table Optimizer selects a floating-point solution as the best solution when all of these conditions are met:

• The floating-point solution requires equal or less memory than a fixed-point solution.

• Both the InputTypes and OutputType properties of the FunctionApproximation.Problem object specify a floating-point data type.

• The WordLengths property of the FunctionApproximation.Options object includes word lengths corresponding to a single-precision or double-precision data type.

Limitations

• Lookup table objects and breakpoint objects are not supported in a model mask workspace.

Algorithms

Required Specifications

Functions and function handles that you approximate must meet the following criteria.

• The function must be time-invariant.

• The function must operate element-wise, meaning for each input there is one output.

• The function must not contain states.

For more information, see Vectorization.

Infinite Upper and Lower Input Bounds

When a Problem object specifies infinite input ranges and the input type is non-floating-point, during the approximation, the software infers upper and lower ranges based on the range of the input data type. The resulting FunctionApproximation.LUTSolution object specifies the bounds that the algorithm used during the approximation, not the originally specified infinite bounds.

Upper and Lower Input Bounds and Input Data Type Range

If the InputLowerBounds or InputUpperBounds specified for a Problem object fall outside the range of the specified InputTypes, the algorithm uses the range of the data type specified by InputTypes for the approximation.

In cases where the BreakpointSpecification property of the FunctionApproximation.Options object is set to 'EvenSpacing', but the InputUpperBounds or InputLowerBounds property of the FunctionApproximation.Problem object is equal to the range of the InputTypes, the algorithm does not attempt to find a solution using 'EvenPow2Spacing'.

Version History

Introduced in R2018a

R2022a: Support for curve fitting objects

The FunctionApproximation.Problem object now supports curve fitting cfit (Curve Fitting Toolbox) objects as valid inputs for approximation.

R2022a: Improved memory reduction for 1D and flat interpolation

The Lookup Table Optimizer has an improved algorithm for lookup table value and breakpoint optimization for one-dimensional functions with flat interpolation. This enhancement can enable improved memory reduction of the optimized lookup table and faster completion of the lookup table optimization process.

This improvement applies when the function to approximate is one-dimensional and all of these options are specified in FunctionApproximation.Options:

• Interpolation is set to Flat.

• BreakpointSpecification is set to ExplicitValues.

• OnCurveTableValues is set to false.

R2021b: Generate an optimized lookup table approximation as a MATLAB function

You can now use the FunctionApproximation.Problem object to generate an optimized lookup table approximation as a MATLAB function. To generate MATLAB function, in a FunctionApproximation.Options object, set the ApproximateSolutionType property to MATLAB.

The generated MATLAB function is editable and supports C/C++ code generation using MATLAB Coder™.

R2021a: Lookup table optimization support for functions with scalar inputs

Previously, the FunctionApproximation.Problem class required that functions and function handles to approximate were vectorized, meaning that for each input, there is exactly one output. Lookup table optimization now fully supports approximation of Simulink blocks and subsystems that only allow scalar inputs.

R2021a: Improved lookup table value optimization

The Lookup Table Optimizer has an improved algorithm for lookup table value optimization for the Flat and Nearest interpolation methods when off-curve table values are allowed. This enhancement can enable faster completion of the lookup table optimization process and improved memory reduction of the optimized lookup table.