System and method of global optimization using artificial neural networks

• US 5,377,307 A
• Filed: 10/07/1992
• Issued: 12/27/1994
• Est. Priority Date: 10/07/1992
• Status: Expired due to Fees

• Current Assignee: Schlumberger Technology Corporation (SLB)
• Original Assignee: Schlumberger Technology Corporation (SLB)
• Inventors: Hoskins, Josiah C., Kramer, Glenn A.
• Primary Examiner(s): MacDonald, Allen R.
• Assistant Examiner(s): Dorvil, Richemond
• Application Number:
  US07/957,565
• Time in Patent Office: 811 Days
• Field of Search: 395/22, 395/23, 364/578, 364/576
• US Class Current: 706/19
• CPC Class Codes: B25J 9/1605 Simulation of manipulator lay-out, design, modelling of manipulator
• CPC Class Codes: G05B 2219/39404 Design of manipulator
• CPC Class Codes: G05B 2219/39405 Develop inverse model of system with ann
• CPC Class Codes: G05B 2219/39406 Obtain optimal parameters of model of system
• CPC Class Codes: G05B 2219/40281 Closed kinematic loop, chain mechanisms, closed linkage systems
• CPC Class Codes: G06F 30/20 Design optimisation, verification or simulation optimisation, verification or simulation of circuit designs G06F30/30
• CPC Class Codes: G06Q 10/04 Forecasting or optimisation specially adapted for administrative or management purposes, e.g. linear programming or "cutting stock problem" market predictions or forecasting for commercial activities G06Q30/0202

Abstract

A method of global optimization of complex, highly nonlinear, multivariant systems is described. An artificial neural network (ANN) is trained to create an approximate inverse model. The desired behavior for a particular system is then input to the inverse model to derive approximate model parameters for the particular system. Optimization of the approximate model parameters yields optimal model parameters. The method is applied to the synthesis of mechanical linkages where examples of a type of linkage mechanism are used to train an ANN and derive the approximate inverse model. Inverse models for a number of linkage mechanism types are derived and stored. For a linkage mechanism with unknown linkage parameters, a power spectrum representation of the coupler curve is developed and the inverse model for the type of linkage mechanism retrieved. The representation of the desired coupler curve is input and the approximate linkage parameters derived. Optimization further refines the linkage parameters.

Claims

1. A method of global optimization comprising the steps of:

a. developing an approximate inverse model of a system using an artificial neural network (ANN), including the substeps

constructing a forward model having inputs and outputs with model parameters as inputs and behaviors as outputs,

building a set of training examples consisting of a set of model parameters and their corresponding behaviors as derived from the forward model,

developing an ANN with said set of training examples to build an inverse model, said inverse model having inputs and outputs with behaviors as inputs and model parameters as outputs

storing said inverse model;

b. determining the approximate model parameters of a system given a desired behavior of the system as inputs including the substeps

retrieving said inverse model, computing the approximate model parameters from said inverse model; and

c. obtaining optimal model parameters for said system given the approximate model parameters including the substeps

initializing a local optimization process with the approximate model parameters,

optimizing said approximate model parameters using the forward model and the optimization process resulting in the optimal model parameters.

2. The method of claim 1, including the steps of:

repeating step a. for multiple systems and storing said inverse model for each system in a computer memory,

selecting, during said retrieving substep of step b. the inverse model in said memory corresponding to a specific system.

3. The method of claim 1, where said system includes constraints, the local optimization process is an iterative process with constraints of the system imposed as penalty functions.

4. The method of claim 3, where the local optimization process is a Levenberg-Marquardt type of optimization.

5. The method of claim 1, where a representation of the desired behavior as a function of space-time is input to the inverse model.

6. The method of claim 5, including the substep of transforming the representation to a function of frequency.

7. The method of claim 6, the representation comprising a Fourier transform.

8. The method of claim 1, said forward model comprising an artificial neural network.

9. The method of claim 1, where the system corresponds to a particular type of mechanical linkage and the optimal model parameters are linkage parameters including link dimensions.

10. The method of claim 1, the developing step a. further comprising:

a. providing a number of examples of a type of linkage mechanism where each example includes linkage parameters and a representation of the linkage operation as a function of time;

b. training said artificial neural network (ANN) using the linkage examples to build said inverse model of the mechanism which uses representation as input and linkage parameters as output;

the determining step b. further comprising;

c. inputting a representation of the desired behavior of the linkage mechanism to the inverse model; and

d. computing approximate linkage parameters from said inverse model.

11. The method of claim 10, including the step of optimizing the approximate linkage parameters to derive optimal linkage parameters.

12. The method of claim 10, including the steps of:

repeating steps a and b for a plurality of types of linkage mechanisms;

storing the corresponding inverse model for each type; and

determining the type of said linkage mechanism and retrieving the corresponding inverse model for use in step c.

13. The method of claim 10, where said linkage operation is a function of space-time, including the substep of transforming the representation to the frequency domain.

14. The method of claim 10, said providing step a. including the substeps for deriving said representation comprising

describing as a coupler curve a path in space of a point on the linkage mechanism as the linkage mechanism is operated,

plotting the coupler curve as a cyclic plot of curvature versus length of curve perimeter, and

deriving a power spectrum of the plot.

15. The method of claim 10, said providing step a. including the substep of constructing a forward model which uses linkage parameters as inputs and a representation of linkage dynamics to build training set examples.