Method and apparatus using a decision tree in an adjunct system cooperating with another physical system

• US 5,481,649 A
• Filed: 03/09/1993
• Issued: 01/02/1996
• Est. Priority Date: 03/09/1993
• Status: Expired due to Fees

• Current Assignee: Birdwell J. Douglas, Roger D. Horn
• Original Assignee: The University Of Tennessee Research Corporation (University of Tennessee)
• Inventors: Horn, Roger D., Birdwell, J. Douglas
• Primary Examiner(s): Downs, Robert W.
• Application Number:
  US08/028,662
• Time in Patent Office: 1,029 Days
• Field of Search: 395/75, 395/77, 395/51, 395/904, 395/906, 364/148
• US Class Current: 706/59
• CPC Class Codes: G05B 13/028 using expert systems only
• CPC Class Codes: Y10S 706/904 Manufacturing or machine, e.g. agricultural machinery, machine tool
• CPC Class Codes: Y10S 706/906 Process plant

Abstract

A method and apparatus for use in an adjunct system, such as a control system, operating in cooperation with a linear or non-linear physical system, such as a motor. The physical system performs a function and produces data that are monitored, and attribute values, which may be continuous valued, and system state information are produced from the data. For example, the attributes may include various motor currents and the system state information may include the motor flux. A controller and a decision tree produce decisions and commands to assist the physical system in performing its function. Certain decisions made by the decision tree are identified as being bad decisions, and the decision tree is adapted to produce a decision tree that has improved performance. Preferably, the adaptation method maintains the size of the adapted decision tree within a constant limit during the adaptation, and the adaptation is repetitively performed to successively produce new currently active decision trees.

Claims

1. A method for use in a control system, said method for monitoring and controlling a linear or non-linear physical system capable of operating in a plurality of physical system states, said physical system performing a function and producing sets of data that include information as to the performance of the function by the physical system, said control system including a monitor for receiving and monitoring the data of the physical system, a currently active decision tree for making decisions;

and a controller for issuing commands for changing the operation of the physical system in accordance with the decisions, the method comprising;

a. collecting data and producing attributes and system state information;

b. using the attributes and system state information to evaluate the currently active decision tree to produce decisions.

c. applying the decisions in the controller to issue commands;

d. transmitting the commands to the physical system to control its operation;

e. repeating the aforementioned steps a-d to repetitively control the operation of the physical system;

f. while repeating steps a-d;

i. identifying decisions made by the currently active decision tree that were bad decisions based on an evaluation of the collected data;

ii. adapting a tree corresponding to the currently active decision tree based on the identified bad decisions to produce an adapted decision tree that has improved performance as compared to the currently active decision tree;

iii. maintaining the size of the adapted decision tree within a constant limit;

iv. using the adapted decision tree for the currently active decision tree to improve performance of the physical system; and

v. repeating the aforementioned steps i-iv to repetitively produce adapted decision trees that are successively used for the currently active decision tree while the size of each successive adapted decision tree is maintained within the constant limit, whereby the currently active decision tree is constantly updated to improve performance.

2. The method of claim 1 wherein said identifying step comprises:

providing a system reference model;

calculating optimal commands for the controller based on the collected data, attributes and system state information and the system reference model; and

using the commands issued by the controller and the calculated optimal commands to identify bad decisions.

3. The method of claim 1 wherein said adapting step comprises:

adapting said decision tree using an incremental adaptation method that modifies portions of the decision tree while maintaining the size of the decision tree within a constant limit.

4. The method of claim 1 wherein said adapting step comprises:

adapting said decision tree using a regeneration method that modifies the decision tree while maintaining the size of the decision tree within a constant limit.

5. A method for use in a control system, said method for monitoring and controlling a linear or non-linear physical system capable of operating in a plurality of physical system states, said physical system performing a function and producing sets of data that include information as to the performance of the function by the physical system, said control system including a monitor for receiving and monitoring the data of the physical system, and a currently active decision tree for making decisions, said currently active decision tree being derived from a training set of attribute values and correct decisions, said control system further including a controller issuing commands for changing the operation of the physical system in accordance with the decisions, the method comprising:

a. collecting data and producing attributes and system state information;

b. using the attributes and system state information to evaluate the currently active decision tree to produce decisions;

c. applying the decisions in the controller to issue commands;

d. transmitting the commands to the physical system to control its operation;

e. repeating the aforementioned steps a-d to repetitively control the operation of the physical system;

f. while repeating the aforementioned steps a-d;

i. providing a system reference model and calculating optimal commands for controlling the physical system based on collected data, attributes and system state information;

ii. comparing the optimal commands to the commands that were actually issued based on the decisions made by the decision tree;

iii. identifying decisions made by the decision tree that were bad decisions based on the comparison of the optimal commands and the actually issued commands;

iv. for each bad decision, identifying corrected decisions based on the optimal command used in said comparing;

v. organizing the corrected decisions and attributes associated with the bad decisions into new examples for a training set;

vi. producing a new training set by adding the new examples to, and removing selected examples from, the training set so that the new training set is maintained at a size below a constant limit;

vii. adapting the currently active decision tree based on the new training set to produce an adapted decision tree that has improved performance as compared to the currently active decision tree and has a size that is maintained within a constant limit;

viii. using the adapted decision tree for the currently active decision tree to improve performance of the control system; and

ix. repeating the aforementioned steps i-viii to repetitively produce adapted decision trees that are successively used for the currently active decision tree while the size of each successive adapted decision tree is maintained within a constant limit, whereby the currently active decision tree is successively updated to improve performance.

6. The method of claim 5 wherein the step of collecting and producing further comprises producing continuous valued attributes.

7. A method for use in a control system, said method for monitoring and controlling a non-linear physical system capable of operating in a plurality of system states, said physical system performing a function and producing data that include information as to the performance of the function by the physical system, said control system including a monitor for receiving and monitoring the data of the physical system;

the method comprising;

a. initially creating a currently active decision tree from examples that include attribute values and decisions, said decision tree including nodes that are associated with examples;

b. collecting data and producing attributes having values and system state information;

c. using the currently active decision tree and a controller to make decisions and issue commands for changing the operation of the physical system in accordance with the decisions;

d. repeating the aforementioned steps a-d to repetitively control the operation of the physical system;

e. while repeating the aforementioned steps a-c;

i. providing a system reference model and calculating optimal commands for controlling the physical system based on the data, attribute values and system state information;

ii. comparing the optimal commands to the commands that were actually issued based on the decisions made by the currently active decision tree;

iii. identifying decisions made by the currently active decision tree that were bad decisions based on the comparison of the optimal commands and the actually issued commands;

iv. for each bad decision, identifying corrected decisions based on the optimal command used in said comparing;

v. organizing the attribute values associated with each identified bad decision into a new example;

vi. including each new example in the examples;

vii. associating each new example with at least one node in the currently active decision tree;

viii. for each new example, identifying an old example for being removed from the examples;

ix. removing the identified old example from the examples and removing all of the associations of the identified old example, so that the steps of including new examples and removing identified old examples maintains the number of examples below a constant limit;

x. adapting the currently active decision tree by incremental adaptation based on the inclusion of new examples and removal of identified old examples to produce an adapted decision tree that has a size that is maintained within a constant limit;

xi. using the adapted decision tree for the currently active decision tree; and

xii. repeating the aforementioned steps i-xi to repetitively produce adapted decision trees that are successively used for the currently active decision tree while the size of each successive adapted decision tree is maintained within a constant limit, whereby the currently active decision tree is constantly updated without growing the adapted decision tree.

8. A method for use in a control system, said method for monitoring and controlling a non-linear physical system capable of operating in a plurality of system states, said physical system performing a function and producing data that include information as to the performance of the function by the physical system, said control system including a monitor for receiving and monitoring the data of the physical system,

the method comprising:

a. collecting data and producing continuous valued attributes having values and system state information;

b. using the currently active decision tree and a controller to make decisions and issue commands for changing the operation of the physical system in accordance with the decisions;

c. repeating the aforementioned steps a-b while;

i. providing a system reference model and calculating optimal commands for controlling the physical system based on the data, continuous valued attribute values and system state information;

ii. comparing the optimal commands to the commands that were actually issued based on the decisions made by the currently active decision tree;

iii. identifying decisions made by the currently active decision tree that were bad decisions based on the comparison of the optimal commands and the actually issued commands;

iv. for each bad decision, identifying corrected decisions based on the optimal command used in said comparing;

v. organizing the attribute values associated with each identified bad decision into a new example;

vi. including each new example in the examples and associating each new example with at least one node in the currently active decision tree;

vii. adapting the currently active decision tree by incremental adaptation based on the inclusion of new examples to produce an adapted decision tree;

viii. using the adapted decision tree for the currently active decision tree; and

ix. repeating the aforementioned steps i-viii to repetitively produce adapted decision trees that are successively used for the currently active decision tree.

9. An improved method for use in a controller system, said method for monitoring and controlling a motor system having a motor capable of operating at a plurality of motor speeds, said motor system producing motor currents and motor current error signals, said motor performing a function and producing continuous valued data that include information as to the performance of the function by the physical system, said system including a monitor for receiving and monitoring the data of the physical system, the method comprising:

initially creating a currently active decision tree from examples that include continuous valued attribute values and decisions, said currently active decision tree including nodes that are associated with examples;

a. collecting data and producing continuous valued attributes and system state information;

b. using the currently active decision tree and a controller to make decisions and issue commands for changing the operation of the physical system in accordance with the decisions;

c. repeating the aforementioned steps a-b while;

i. providing a system reference model and calculating optimal commands for controlling the motor system based on the data, attribute values and system state information;

ii. comparing the optimal commands to the commands that were actually issued based on the decisions made by the currently active decision tree;

iii. identifying decisions made by the currently active decision tree that were bad decisions based on the comparison of the optimal commands and the actually issued commands;

iv. for each bad decision, identifying corrected decisions based on the optimal command used in said comparing;

v. organizing the optimal decision and the attribute values associated with each bad decision into a new example in the format of the training set;

vi. including each new example in the examples; and

vii. associating each new example with at least one node in the currently active decision tree;

viii. for each new example, identifying an old example for being removed from training set;

ix. removing the identified old example from the examples so that the steps of including new examples and removing identified old examples maintains the number of examples below a constant limit;

x. adapting the currently active decision tree by incremental adaptation based on the inclusion of new examples and the removal of identified old examples;

xi. using the adapted decision tree for the currently active decision tree; and

xii. repeating the aforementioned steps i-xi to repetitively produce adapted decision trees that are successively used for the currently active decision tree while the size of each successive adapted decision tree is maintained within a constant limit, whereby the currently active decision tree is constantly updated to improve performance without growing the adapted decision tree to a size that would impede its performance or exceed the capacity of the storage device in which the currently active decision tree is located.

10. A control system for monitoring and controlling a linear or non-linear physical system capable of operating in a plurality of physical system states, said physical system performing a function and repetitively producing data that include information as to the performance of the function by the physical system that may change over time, said control system comprising:

a monitor for receiving and monitoring the repetitively produced data of the physical system;

a currently active decision tree implemented in a device for repetitively making decisions based on the received data; and

a controller for issuing commands for changing the operation of the physical system based on the decisions;

a processing system that operates, while the currently active decision tree and the controller issue commands to change the operation of the physical system, for;

a. collecting data from the monitor and produceing attribute values and system state information;

b. identifying decisions made by the currently active decision tree that were bad decisions based on an evaluation of the attribute values and system state information;

c. for each bad decision, identifying corrected decisions;

d. adapting the currently active decision tree based on the identified bad decisions and identified corrected decisions to produce an adapted decision tree that has improved performance as compared to the currently active decision tree, while maintaining the size of the adapted decision tree within a constant limit;

e. using the adapted decision tree for the currently active decision tree to improve performance of the controller system; and

f. repeating the aforementioned functions a-e to repetitively produce adapted decision trees that are successively used for the currently active decision tree while the size of each successive adapted decision tree is maintained within a constant limit.

11. The control system of claim 10 wherein said device is a programmable gate array and said currently active decision tree is implemented by said programmable gate array.

12. The control system of claim 10 wherein said device comprises a programmable gate array for implementing a decision tree evaluation function and memory connected with said programmable gate array for storing at least one said decision tree.

13. The control system of claim 10 wherein:

said device comprises at least one programmable gate array and at least two decision trees are implemented by said device, one of which is the currently active decision tree; and

said processing system is operable to switch and designate to another decision tree as the currently active decision tree.

14. The control system of claim 10 wherein said device comprises two programmable gate arrays and one decision tree is implemented by each said array.

15. A method for producing and incrementally adapting a decision tree comprising:

providing examples including attribute values and a decision;

initially creating a decision tree having at least one node and being consistent with the examples;

associating the examples and nodes of the decision tree;

providing a new example; and

adapting the tree, at least in part, by including the new example in the examples and associating the new example with at least one node in the decision tree to produce an adapted tree.

16. The method of claim 15 wherein said attributes have continuous values.

17. The method of claim 15 further comprising:

identifying an example for being removed from the examples; and

removing the identified example and the associations between the nodes and the identified example.

18. The method of claim 17 wherein said attributes have continuous values.