Method and system for constructing cognitive programs

• US 7,437,335 B2
• Filed: 11/23/2005
• Issued: 10/14/2008
• Est. Priority Date: 12/07/2004
• Status: Active Grant

• Current Assignee: Eric Baum
• Original Assignee: Eric Baum
• Inventors: Baum, Eric
• Primary Examiner(s): Hirl; Joseph P
• Application Number:
  US11/285,937
• Publication Number:
  US 20060184916A1
• Time in Patent Office: 1,056 Days
• Field of Search: 706/12, 706/14, 706/45, 717/106, 395/500
• US Class Current: 706/12
• CPC Class Codes: G06F 9/30145 Instruction analysis, e.g. decoding, instruction word fields
• CPC Class Codes: G06N 20/00 Machine learning

Abstract

The present invention is directed to a method of creating a program to solve a problem pertaining to a first task. In the method, user input data pertaining to a second task is received from an operator. The second task is different from the first task. A number of subprograms are automatically obtained based on the received user input data. The program is then created based on the obtained subprogram or subprograms.

Claims

1. A method of creating a program embodied in a computer readable memory to cause a computer to perform a first task of achieving a first constraint or a first set of constraints when run on an input or inputs taken from a class of inputs, the method comprising:

performing two or more times an operation comprising,

(i) receiving input data pertaining to a respective other task of achieving a respective other constraint or other set of constraints when run on inputs taken from a respective other class of inputs, in which the respective other constraint or other set of constraints is different from the first constraint or the first set of constraints, and different from the previous respective other constraints or other sets of constraints, and in which said input data includes a number of examples of the respective other task in which each example is a particular input in the class of inputs for the respective other task, and in which said input data further includes information pertaining to a respective fitness function which indicates how well a respective example or examples are solved for the respective other task and in which said respective fitness function is different from the respective fitness function for each previous other task;

(ii) obtaining a respective instruction set having a respective collection of instruction(s), function(s), procedure(s) and/or module(s);

(iii) automatically obtaining a respective subprogram by learning from the number of examples in the received input data that is acceptable in accordance with the respective fitness function for the respective other task, in which the learning for obtaining a respective subprogram includes obtaining two candidate subprograms each different from the other and each composed from the instruction(s), function(s), procedure(s), and/or module(s) in the respective instruction set, determining how well each of the two candidate subprograms perform the respective other task when run on one or more of the number of examples of the respective other task by use of the respective fitness function associated with the respective other task, in which the respective fitness function is different from each previous fitness function, determining which one of the two candidate subprograms performs better according to the respective fitness function, determining if the better candidate subprogram performs acceptably and if the better candidate subprogram is determined to perform acceptably selecting the better candidate subprogram as an obtained subprogram;

(iv) adding the respective obtained subprogram to a storage library and associating a respective name operable to invoke the stored respective program, and adding the name to previously obtained instruction(s), function(s), procedure(s), and/or module(s) available for inclusion in the respective instruction set or sets; and

obtaining another instruction set having a respective collection of instruction(s), function(s), procedure(s), and/or module(s) including one or more of the respective names, in which the another instruction set is different from each of the previously obtained respective instruction sets; and

creating said program to perform said first task composed from the instruction(s), function(s), procedure(s), and/or module(s) in the another instruction set so that the program contains the name or names of one or more of the respective automatically obtained subprograms each having been determined to perform one of the other tasks acceptably which is different from the first task by use of the respective fitness function associated with the respective other task when run on one or more of the number of examples.

2. The method according to claim 1, in which the respective instruction set obtained one time the operation is performed is different from the respective instruction set obtained a subsequent time the operation is performed.

3. The method according to claim 2 in which one or more of the instruction(s), function(s), procedure(s) and/or module(s) in a first respective instruction set obtained the one time the operation is performed are not contained in a second respective instruction set obtained the subsequent time the operation is performed.

4. The method according to 1, in which the creating said program to perform said first task composed from the instruction(s), function(s), procedure(s), and/or module(s) in the another instruction set includes receiving first task input data pertaining to the first task of achieving the first constraint or the first set of constraints when run on the input or inputs taken from the class of inputs, in which said first task input data includes a number of examples of the first task in which each first task example is a particular input in the class of inputs for the first task, and in which said first task input data further includes information pertaining to a first task fitness function which indicates how well an example or examples are solved for the first task;

and automatically obtaining said program by learning from the number of examples in the received first task input data that is acceptable in accordance with the first fitness function, in which the learning for automatically obtaining said program includes obtaining two candidate programs each different from the other and each composed from the instruction(s), function(s), procedure(s), and/or module(s) in the another instruction set, determining how well each of the two candidate programs perform the first task when run on one or more of the number of examples of the first task by use of the first task fitness function, determining which one of the two candidate programs performs better according to the first task fitness function, determining if the better candidate program performs acceptably and if the better candidate program is determined to perform acceptably selecting the better candidate program as the created said program.

5. The method according to 1, in which the creating said program to perform said first task composed from the instruction(s), function(s), procedure(s), and/or module(s) in the another instruction set includes obtaining a respective procedure having a number of arguments and respectively substituting for one or more of the number of arguments a respective one of the respective automatically obtained subprograms or the respective name thereof.

6. The method according to 5, in which the respective procedure further includes a number of annotations associated with one or more of the number of arguments, and in which the receiving input data pertaining to the respective other task includes using the annotation or annotations associated with a respective argument to obtain the received input data pertaining to the respective other task.

7. The method according to 1, in which one or more of the respective automatically obtained subprograms and/or one or more of the instruction(s), function(s), procedure(s) and/or module(s) in one of the respective instruction sets or the another instruction set includes a respective plurality of agents and an auction mechanism for applying said agents in a coordinated fashion.

8. The method according to 1, further comprising if the better candidate subprogram is determined not to perform acceptably, obtaining another candidate subprogram and determining how well said another candidate subprogram performs and continuing until an acceptable subprogram is obtained or until a predetermined time is reached.

9. A system of creating a program embodied in a computer readable memory to cause a computer to perform a first task of achieving a first constraint or a first set of constraints when run on an input or inputs taken from a class of inputs, the system comprising:

a device which when operating performs two or more times an operation comprising,

(i) receiving input data pertaining to a respective other task of achieving a respective other constraint or other set of constraints when run on inputs taken from a respective other class of inputs, in which the respective other constraint or other set of constraints is different from the first constraint or the first set of constraints, and different from the previous respective other constraints or other sets of constraints, and in which said input data includes a number of examples of the respective other task in which each example is a particular input in the class of inputs for the respective other task, and in which said input data further includes information pertaining to a respective fitness function which indicates how well a respective example or examples are solved for the respective other task and in which said respective fitness function is different from the respective fitness function for each previous other task;

(ii) obtaining a respective instruction set having a respective collection of instruction(s), function(s), procedure(s) and/or module(s);

(iii) automatically obtaining a respective subprogram by learning from the number of examples in the received input data that is acceptable in accordance with the respective fitness function for the respective other task, in which the learning for obtaining a respective subprogram includes obtaining two candidate subprograms each different from the other and each composed from the instruction(s), function(s), procedure(s), and/or module(s) in the respective instruction set, determining how well each of the two candidate subprograms perform the respective other task when run on one or more of the number of examples of the respective other task by use of the respective fitness function associated with the respective other task, in which the respective fitness function is different from each previous fitness function, determining which one of the two candidate subprograms performs better according to the respective fitness function, determining if the better candidate subprogram performs acceptably and if the better candidate subprogram is determined to perform acceptably selecting the better candidate subprogram as an obtained subprogram;

(iv) adding the respective obtained subprogram to a storage library and associating a respective name operable to invoke the stored respective program, and adding the name to previously obtained instruction(s), function(s), procedure(s), and/or module(s) available for inclusion in the respective instruction set or sets; and

a device which when operating obtains another instruction set having a respective collection of instruction(s), function(s), procedure(s), and/or module(s) including one or more of the respective names, in which the another instruction set is different from each of the previously obtained respective instruction sets; and

a device which when operating creates said program to perform said first task composed from the instruction(s), function(s), procedure(s), and/or module(s) in the another instruction set so that the program contains the name or names of one or more of the respective automatically obtained subprograms each having been determined to perform one of the other tasks acceptably which is different from the first task by use of the respective fitness function associated with the respective other task when run on one or more of the number of examples.

10. The system according to claim 9, in which the respective instruction set obtained one time the operation is performed is different from the respective instruction set obtained a subsequent time the operation is performed.

11. The system according to claim 10, in which one or more of the instruction(s), function(s), procedure(s) and/or module(s) in a first respective instruction set obtained the one time the operation is performed are not contained in a second respective instruction set obtained the subsequent time the operation is performed.

12. The system according to 9, in which creating performed by the device which when operating creates said program to perform said first task composed from the instruction(s), function(s), procedure(s), and/or module(s) in the another instruction set includes receiving first task input data pertaining to the first task of achieving the first constraint or the first set of constraints when run on the input or inputs taken from the class of inputs, in which said first task input data includes a number of examples of the first task in which each first task example is a particular input in the class of inputs for the first task, and in which said first task input data further includes information pertaining to a first task fitness function which indicates how well an example or examples are solved for the first task;

and automatically obtaining said program by learning from the number of examples in the received first task input data that is acceptable in accordance with the first fitness function, in which the learning for automatically obtaining said program includes obtaining two candidate programs each different from the other and each composed from the instruction(s), function(s), procedure(s), and/or module(s) in the another instruction set, determining how well each of the two candidate programs perform the first task when run on one or more of the number of examples of the first task by use of the first task fitness function, determining which one of the two candidate programs performs better according to the first task fitness function, determining if the better candidate program performs acceptably and if the better candidate program is determined to perform acceptably selecting the better candidate program as the created said program.

13. The system according to 9, in which creating performed by the device which when operating creates said program to perform said first task composed from the instruction(s), function(s), procedure(s), and/or module(s) in the another instruction set includes obtaining a respective procedure having a number of arguments and respectively substituting for one or more of the number of arguments a respective one of the respective automatically obtained subprograms or the respective name thereof.

14. The system according to 13, in which the respective procedure further includes a number of annotations associated with one or more of the number of arguments, and in which the receiving input data pertaining to the respective other task includes using the annotation or annotations associated with a respective argument to obtain the received input data pertaining to the respective other task.

15. The system according to 9, in which one or more of the respective automatically obtained subprograms and/or one or more of the instruction(s), function(s), procedure(s) and/or module(s) in one of the respective instruction sets or the another instruction set includes a respective plurality of agents and an auction mechanism for applying said agents in a coordinated fashion.

16. The system according to 9, further comprising a device which when operating, if the better candidate subprogram is determined not to perform acceptably, obtains another candidate subprogram and determines how well said another candidate subprogram performs and continues until an acceptable subprogram is obtained or until a predetermined time is reached.

17. A system of creating a program embodied in a computer readable memory to cause a computer to perform a first task of achieving a first constraint or a first set of constraints when run on an input or inputs taken from a class of inputs, the system comprising:

means for performing two or more times an operation comprising,

(i) receiving input data pertaining to a respective other task of achieving a respective other constraint or other set of constraints when run on inputs taken from a respective other class of inputs, in which the respective other constraint or other set of constraints is different from the first constraint or the first set of constraints, and different from the previous respective other constraints or other sets of constraints, and in which said input data includes a number of examples of the respective other task in which each example is a particular input in the class of inputs for the respective other task, and in which said input data further includes information pertaining to a respective fitness function which indicates how well a respective example or examples are solved for the respective other task and in which said respective fitness function is different from the respective fitness function for each previous other task;

(ii) obtaining a respective instruction set having a respective collection of instruction(s), function(s), procedure(s) and/or module(s);

(iii) automatically obtaining a respective subprogram by learning from the number of examples in the received input data that is acceptable in accordance with the respective fitness function for the respective other task, in which the learning for obtaining a respective subprogram includes obtaining two candidate subprograms each different from the other and each composed from the instruction(s), function(s), procedure(s), and/or module(s) in the respective instruction set, determining how well each of the two candidate subprograms perform the respective other task when run on one or more of the number of examples of the respective other task by use of the respective fitness function associated with the respective other task, in which the respective fitness function is different from each previous fitness function, determining which one of the two candidate subprograms performs better according to the respective fitness function, determining if the better candidate subprogram performs acceptably and if the better candidate subprogram is determined to perform acceptably selecting the better candidate subprogram as an obtained subprogram;

(iv) adding the respective obtained subprogram to a storage library and associating a respective name operable to invoke the stored respective program, and adding the name to previously obtained instruction(s), function(s), procedure(s), and/or module(s) available for inclusion in the respective instruction set or sets; and

means for obtaining another instruction set having a respective collection of instruction(s), function(s), procedure(s), and/or module(s) including one or more of the respective names, in which the another instruction set is different from each of the previously obtained respective instruction sets; and

means for creating said program to perform said first task composed from the instruction(s), function(s), procedure(s), and/or module(s) in the another instruction set so that the program contains the name or names of one or more of the respective automatically obtained subprograms each having been determined to perform one of the other tasks acceptably which is different from the first task by use of the respective fitness function associated with the respective other task when run on one or more of the number of examples.

18. The system according to claim 17, in which the respective instruction set obtained one time the operation is performed is different from the respective instruction set obtained a subsequent time the operation is performed.

19. The system according to claim 18, in which one or more of the instruction(s), function(s), procedure(s) and/or module(s) in a first respective instruction set obtained the one time the operation is performed are not contained in a second respective instruction set obtained the subsequent time the operation is performed.

20. The system according to 17, in which creating performed by the means for creating said program includes receiving first task input data pertaining to the first task of achieving the first constraint or the first set of constraints when run on the input or inputs taken from the class of inputs, in which said first task input data includes a number of examples of the first task in which each first task example is a particular input in the class of inputs for the first task, and in which said first task input data further includes information pertaining to a first task fitness function which indicates how well an example or examples are solved for the first task;

and automatically obtaining said program by learning from the number of examples in the received first task input data that is acceptable in accordance with the first fitness function, in which the learning for automatically obtaining said program includes obtaining two candidate programs each different from the other and each composed from the instruction(s), function(s), procedure(s), and/or module(s) in the another instruction set, determining how well each of the two candidate programs perform the first task when run on one or more of the number of examples of the first task by use of the first task fitness function, determining which one of the two candidate programs performs better according to the first task fitness function, determining if the better candidate program performs acceptably and if the better candidate program is determined to perform acceptably selecting the better candidate program as the created said program.