Distributed machine learning intelligence development systems

• US 9,390,376 B2
• Filed: 10/15/2013
• Issued: 07/12/2016
• Est. Priority Date: 10/15/2013
• Status: Active Grant

• Current Assignee: Lockheed Martin Corporation (Martin Marietta Corporation)
• Original Assignee: Lockheed Martin Corporation (Martin Marietta Corporation)
• Inventors: Brant, Jonathan Charles, Harrison, Gregory A., Worden, Eric W., Smith, David A.
• Primary Examiner(s): Chaki, Kakali
• Assistant Examiner(s): Wu, Fuming
• Application Number:
  US14/053,811
• Publication Number:
  US 20150106308A1
• Time in Patent Office: 1,001 Days
• Field of Search: None
• US Class Current: 1/1
• CPC Class Codes: G06N 3/126 Evolutionary algorithms, e.g. genetic algorithms or genetic programming
• CPC Class Codes: G06N 5/043 Distributed expert systems; Blackboards

Abstract

A system, method, and computer-readable instructions for a distributed machine learning system are provided. A plurality of distributed learning environments are in communication over a network, wherein each environment has a computing device having a memory and a processor coupled to the memory, the processor adapted implement a learning environment via one or more agents in a rules-based system, wherein the agents learn to perform tasks in their respective learning environment; and a persistent storage in which knowledge comprising a plurality of rules developed by the agents for performing the tasks are stored, wherein the knowledge is tagged and shared with other agents throughout the plurality of distributed learning environments.

Claims

1. A distributed machine learning system, comprising:

a plurality of distributed learning environments in communication over a network, wherein each environment comprises;

a computing device having a memory and a processor coupled to the memory, the processor adapted to implement a learning environment via one or more agents in a rules-based system, wherein the agents learn to perform tasks in their respective learning environment to develop knowledge;

a storage in which knowledge comprising a plurality of rule sets developed by the agents for performing the tasks are stored, wherein the knowledge is tagged to facilitate identification, storage and retrieval;

an ontology that is weighted to provide, in response to at least one request by the one or more agents, classification of at least a close choice of knowledge to share rule sets developed by the agents in the plurality of distributed learning environments for implementing tasks which are related, but not necessarily the same so that matched parts from the at least close choice of the knowledge provided from at least one sending agent is included with an unmatched part of the rule sets of at least one requesting agent, wherein the at least close choice of knowledge comprising a highest score indicating a level of match of the capabilities between a desired and an available rules lists and providing classification comprises tagging the knowledge with words that define what it does using an ontology-based tagging system;

an interface for sharing at least one of the tagged knowledge and the at least close choice of knowledge with other agents throughout the plurality of distributed learning environments when the at least one requesting agent requests at least one of the tagged knowledge and the close choice of knowledge.

2. The distributed machine learning system of claim 1, wherein the processor is a genetic processor adapted to implement the learning environment to provide a genetically evolving set of rules for each of the agents.

3. The distributed machine learning system of claim 1, wherein the learning environment is a Complex Adaptive System.

4. The distributed machine learning system of claim 1, further comprising a simulator to provide simulations to facilitate learning in the learning environment.

5. The distributed machine learning system of claim 1, wherein the plurality of distributed learning environments comprise one or more duplicate learning environments for parallel learning of tasks, wherein the rule sets developed by the agents for implementing the tasks that are substantially the same are adapted to be directly shareable to other agents in the one or more duplicate learning environments.

6. The distributed machine learning system of claim 1, wherein the matched parts from the at least close choice of the knowledge is included with an unmatched part of the rule sets of the at least one requesting agent to provide for a new rule sets.

7. The distributed machine learning system of claim 1, wherein the knowledge is also metatagged with one or more higher levels in an ontology of behavior.

8. The distributed machine learning system of claim 1 wherein the processor is adapted to provide an ontological abstraction of the classifications of the knowledge encoded in the rules that is at least a close choice of the knowledge sought.

9. The distributed machine learning system of claim 1, wherein the processor is adapted to provide an abstraction of targets of actions in the rules.

10. The distributed machine learning system of claim 1, wherein at least some of the agents are adapted to actively learn by receiving rewards or punishments as they complete a task with a goal to increase rewards and minimize punishments.

11. The distributed machine learning system of claim 1, further comprising a Best Rule List (BRL) stored for a set of learned rules that performs better than a prior set for a certain task.

12. The distributed machine learning system of claim 11, wherein at least some of the agents are adapted to execute the Best Rule List and explore the learning environment in an attempt to find a better Best Rule List.

13. The distributed machine learning system of claim 12, wherein an agent learns the environment by sending out a query to other agents in the distributed learning environments requesting any rule lists that deal with a task or a part of a task.

14. The distributed machine learning system of claim 1, further comprising a talent scout application adapted to use a task description for a desired task and the tagged knowledge to find partial or whole rule lists that are suitable for the desired task.

15. The distributed machine learning system of claim 14 wherein the talent scout application is adapted to generalize tags to find partial or whole rule lists that are suitable for the desired task.

16. The distributed machine learning system of claim 1, wherein the processor is adapted to use higher ontological levels of match with less specificity to find partial or whole rule lists that are suitable for a desired task.

17. The distributed machine learning system of claim 1, wherein the knowledge is stored using a first-order-predicate logic FOPL.

18. The distributed machine learning system of claim 1, wherein learning and operation of the agents are substantially controlled by an artificial economy and virtual funds.

19. A method for implementing a distributed machine learning system across a plurality of distributed learning environments in communication over a network, comprising:

implementing, via a processor of a computing device, a learning environment via one or more agents in a rules-based system, wherein the agents learn to perform tasks in their respective learning environment to develop knowledge;

tagging and storing knowledge comprising a plurality of rule sets developed by the agents for performing the tasks; and

providing, in response to at least one request by at least one requesting agent, classification of at least a close choice of knowledge with a weighted ontology to share rule sets developed by the agents in the plurality of distributed learning environments for implementing tasks which are related, but not necessarily the same so that matched parts from the at least close choice of the knowledge provided from at least one sending agent is included with an unmatched part of the rule sets of the at least one requesting agent wherein the at least close choice of knowledge comprising a highest score indicating a level of match of the capabilities between desired and available rules lists, and providing classification comprises;

tagging the knowledge with words that define what it does using an ontology-based tagging system;

providing access to the at least one of tagged knowledge and the close choice of knowledge to allow for sharing with other agents throughout the plurality of distributed learning environments when the at least one requesting agent requests at least one of the tagged knowledge and the close choice of knowledge.

20. A non-transitory computer readable medium containing instructions for implementing a distributed machine learning system across a plurality of distributed learning environments in communication over a network, the instructions when executed cause a processor to:

implement a learning environment via one or more agents in a rules-based system, wherein the agents learn to perform tasks in their respective learning environment to develop knowledge;

tag and store knowledge comprising a plurality of rule sets developed by the agents for performing the tasks;

provide, in response to at least one request by at least one requesting agent, classification of at least a close choice of knowledge with a weighted ontology to share rule sets developed by the agents in the plurality of distributed learning environments for implementing tasks which are related, but not necessarily the same so that matched parts from the at least close choice of the knowledge provided from at least one sending agent is included with an unmatched part of the rule sets of the at least one requesting agent, wherein the at least close choice of knowledge comprising a highest score indicating a level of match of the capabilities between desired and available rules lists, and provide classification comprises;

tag the knowledge with words that define what it does using an ontology-based tagging system; and

provide access to the at least one of tagged knowledge and the close choice of knowledge to allow for sharing with other agents throughout the plurality of distributed learning environments when the at least one requesting agent requests at least one of the tagged knowledge and the close choice of knowledge.