Reluctant Episodic Memory (REM) to Store Experiences of Everyday Interaction With Objects

• US 20080294590A1
• Filed: 09/21/2007
• Published: 11/27/2008
• Est. Priority Date: 05/23/2007
• Status: Active Grant

• Current Assignee: Honda Motor Co., Ltd. (Honda Motor Company)
• Original Assignee: Honda Motor Co., Ltd. (Honda Motor Company)
• Inventors: Gupta, Rakesh, Hennacy, Ken
• Granted Patent: US 7,974,938 B2
• Time in Patent Office: Days
• US Class Current: 706/59
• CPC Class Codes: G06N 3/02 Neural networks
• CPC Class Codes: G06N 5/02 Knowledge representation; Symbolic representation

Abstract

A method and system for storing episodic sequences (events and actions). The system learns episodic sequencing by observing real-world events and actions or by receiving fact data from a database storing common sense facts. The episodic sequences are classified into events and actions, processed to indicate correlations and causality between the events and actions, and generated into linked graphs. The linked graphs may then be used to draw inferences, recognize patterns, and make decisions.

Claims

1. A method of storing and using a set of observations of real-world events and actions, the method comprising:

generating clusters based on the set of observations, each cluster representing correlations between events and actions in a subset of the observations in the set of observations; and

processing the clusters into a first linked graph based on a set of rules to identify correlations between the events and actions in the set of the observations, the first linked graph tagged to indicate the correlations between the events and the actions based on the set of rules.

2. The method of claim 1, further comprising:

storing the clusters in a storage device; and

outputting an effector signal to an effector responsive to receiving a target event signal intended to cause a real-world event, the effector signal generated based on the first linked graph.

3. The method of claim 1, further comprising:

sensing the events and actions using a sensor; and

classifying and formatting the sensed events and actions into a datastream.

4. The method of claim 1, further comprising receiving from a database fact data structured in natural language to represent correlations between the events and actions.

5. The method of claim 1, wherein each cluster comprises a decision point representing a state where an action can be started, and processing the clusters comprises merging two or more decision points of two or more clusters into one decision point to resolve inconsistencies in the two or more clusters.

6. The method of claim 5, wherein each cluster further comprises an edge associated with the decision point to represent a dynamic variable that changes responsive to performing the action.

7. The method of claim 5, wherein each cluster further comprises an edge associated with the decision point to represent the action that is performed in a time frame where a change in a dynamic variable associated with the decision point is observed.

8. The method of claim 7, wherein generating clusters comprises generating a first cluster and a second cluster responsive to detecting a change in the dynamic variable from a first state to a second state, the first cluster comprising a first edge representing a first action performed in a time frame where the dynamic variable is changed from the second state to the first state, and the second cluster comprising a second edge representing a second action performed in a time frame where the dynamic variable is changed from the first state to the second state.

9. The method of claim 1, wherein the set of rules comprises assuming that an event is reversible unless an observation contradicting reversibility of event is obtained.

10. The method of claim 1, further comprising generating a second linked graph for another set of observations by modifying or expanding the first linked graph.

11. The method of claim 1, wherein the subset of observations include one or two observations of the events and actions.

12. A system for storing and using a set of observations of real-world events and actions, the system comprising:

a cluster generator for generating clusters based on the set of observations, each cluster representing correlations between events and actions in a subset of observations in the set of observations; and

a graph processor coupled to the cluster generator for processing the clusters into a first linked graph based on a set of rules to identify correlations between the events and actions, the first linked graph tagged to indicate the correlations between the events and the actions.

13. The system of claim 12, further comprising:

a storage device coupled between the cluster generator and the graph processor for storing the clusters generated by the cluster generator;

an interface module coupled to the graph processor to provide a target event signal intended to cause a real-world event; and

a controller for controlling an effector according to commands generated by the graph processor based on the first linked graph to cause the intended real-world event.

14. The system of claim 12, further comprising:

a sensor for sensing real-world events and actions to generate the sensor signals; and

an encoder coupled between the sensor and the cluster generator for formatting the sensor signals into a datastream identifying the events and the actions.

15. The system of claim 14, further comprising a database storing fact data representing correlation between the events and actions causing the real-world events in natural language, the database associated with the graph generator to provide the fact data to the graph generator.

16. The system of claim 12, wherein each cluster comprises a decision point representing a state where an action can be started, and processing the clusters comprises merging two or more decision points of two or more clusters into one decision point to resolve inconsistencies in the two or more clusters.

17. The system of claim 16, wherein each cluster further comprises an edge associated with the decision point to represent a dynamic variable that changes responsive to performing the action.

18. The system of claim 17, wherein each cluster further comprises an edge associated with the decision point to represent the action that is performed in a time frame where a change in a dynamic variable associated with the decision point is observed.

19. The system of claim 18, wherein the cluster generator generates a first cluster and a second cluster responsive to detecting a change in the dynamic variable from a first state to a second state, the first cluster comprising a first edge representing a first action performed in a time frame where the dynamic variable is changed from the second state to the first state, and the second cluster comprising a second edge representing a second action performed in a time frame where the dynamic variable is changed from the first state to the second state.

20. The system of claim 12, wherein the cluster generator presupposes an event as reversible unless an observation contradicting reversibility of the event is obtained.

21. The system of claim 20, wherein the graph processor tags an edge associated with the irreversible event as having uncertain correlations with the actions that are irreversible.

22. The system of claim 12, wherein the graph processor generates a second linked graph for another set of observations by modifying or expanding the first linked graph.

23. The system of claim 12, wherein the subset of observations include one or two observations of the events and actions.

24. A computer program product comprising a computer readable medium structured to store instructions executable by a processor, the instructions, when executed cause the processor to:

generate clusters based on the set of observations, each cluster representing correlations between events and actions in a subset of the observations in the set of observations; and

process the clusters into a first linked graph based on a set of rules to identify correlations between the events and actions in the set of the observations, the first linked graph tagged to indicate the correlations between the events and the actions based on the set of rules.

25. The computer program product of claim 24, wherein the subset of observations include one or two observations of the events and actions.