Multiple-agent hybrid control architecture for intelligent real-time control of distributed nonlinear processes

• US 5,963,447 A
• Filed: 08/22/1997
• Issued: 10/05/1999
• Est. Priority Date: 08/22/1997
• Status: Expired due to Term

• Current Assignee: Clearsight Systems, Inc.
• Original Assignee: Clearsight Systems, Inc.
• Inventors: Kohn, Wolf, Nerode, Anil
• Primary Examiner(s): Grant, William
• Assistant Examiner(s): Marc, McDieunel, Marc, McDieunel
• Application Number:
  US08/916,418
• Time in Patent Office: 774 Days
• Field of Search: 364/148.04, 364/131, 364/188, 364/474.23, 706/10, 706/46, 706/59, 706/61, 706/45, 395/11, 395/51, 370/276, 702/182, 342/465
• US Class Current: 700/49
• CPC Class Codes: G05B 19/41865 characterised by job scheduling, process planning, material flow
• CPC Class Codes: G05B 2219/33055 Holon, agent executes task and cooperates with other, distributed control
• CPC Class Codes: G05B 2219/33065 Ontogenetic learning, agent learns and adapt its own behaviour
• CPC Class Codes: G05B 2219/33073 Ion control agent has communication, database, suggestion, decision, action, detect
• CPC Class Codes: G06N 5/043 Distributed expert systems; Blackboards
• CPC Class Codes: Y02P 90/02 Total factory control, e.g. smart factories, flexible manufacturing systems [FMS] or integrated manufacturing systems [IMS]

Abstract

A Multiple-Agent Hybrid Control Architecture (MAHCA) uses agents to analyze design, and implement intelligent control of distributed processes. A network of agents can be configured to control more complex distributed processes. The network of agents interact to create an emergent global behavior. Global behavior is emergent from individual agent behaviors and is achieved without central control through the imposition of global constraints on the network of individual agent behaviors. Agent synchronization can be achieved by satisfaction of an interagent invariance principle. At each update time, the active plan of each of the agents in the network encodes equivalent behavior modulo a congruence relation determined by the knowledge clauses in each agents'"'"'s knowledge base. The Control Loop and the Reactive Learning Loop of each agent can be implemented separately. This separation results in an implementation runs faster and with less memory requirements than an unseparated arrangement. A Direct Memory Map (DMM) is to implement the agent architecture. The DMM is a procedure for transforming knowledge and acts as a compiler of agent knowledge by providing a data structure called memory patches, which are used to organize the knowledge contained in each agent'"'"'s Knowledge Base. Content addressable memory is used as the basic mechanism of the memory patch structure. Content addressable memory uses a specialized register called the comparand to store a pattern that is compared with contents of the memory cells. The DMM has two comparands, the Present State Comparand and the Goal Comparand. The MAHCA can be used for compression/decompression for processing and storage of audio or video data.

Claims

1. An agent for controlling a local process of a distributed hybrid system comprising:

a control loop for controlling the local process according to stored requirements that are valid;

a reactive learning loop for modifying the stored requirements when the requirements are not valid; and

wherein the control loop is separated from the reactive learning loop such that while the stored requirements are valid, the control loop operates independently from the reactive learning loop.

2. The agent of claim 1, wherein the control loop comprises:

a knowledge model for storing the requirements as a plan; and

a control inferencer that comprises an abstract state machine that navigates in a domain defined by the knowledge model and determines whether the plan is compatible with a current state of the knowledge model.

3. The agent of claim 2, wherein the plan is fixed until the control inferencer indicates that the plan is not compatible with the current state of the knowledge model.

4. The agent of claim 3, further comprising:

a coupling channel that connects the control loop with the reactive learning loop;

wherein failure terms of the incompatibility of the plan are sent from the control inferencer to the reactive learning loop through the coupling channel, and the coupling channel is active only when the plan is incompatible.

5. The agent of claim 1, wherein the reactive learning loop comprises:

a planner for generating a plan representing the requirements of desired behavior of the system as logic expressions;

a reactive inferencer for determining whether the plan is currently active in the control loop; and

an adapter connected to the inferencer for computing correction terms for the plan if the plan is not currently active in the control loop.

6. The agent of claim 5, further comprising:

a coupling channel that connects the control loop with the reactive learning loop;

wherein a corrected plan is sent from the planner to the control loop through the coupling channel, and the coupling channel is active only when the plan is inactive.

7. The agent of claim 1,

wherein the control loop comprises:

a knowledge model for storing the requirements as a plan; and

a control inferencer that comprises an abstract state machine that navigates in a domain defined by the knowledge model and determines whether the plan is compatible with a current state of the knowledge model; and

wherein the reactive learning loop comprises;

a planner for generating the plan representing the requirements of desired behavior of the system as logic expressions;

a reactive inferencer for determining whether the plan is currently active in the control loop by monitoring the control loop inferencer; and

an adapter connected to the inferencer for computing correction terms for the plan if the plan is not currently active in the control loop.

8. The agent of claim 7, wherein the control loop is connected to the reactive learning loop through first and second coupling channels, wherein

the first coupling channel relays failure terms defining the incompatibility of the plan from the control inferencer to the adapter;

and

the second coupling channel relays the correction terms for the plan from the planner to the control inferencer.

9. The agent of claim 8, wherein

the coupling channels only relay the failure terms and the correction terms when the plan in the control loop is inactive because it is incompatible.

10. The agent of claim 8, wherein the agent is used for data compression from a data source.

11. The agent of claim 10, wherein the data source is a video source.

12. The agent of claim 10, wherein the data source is an audio source.

13. The agent of claim 10, wherein

each data page from the data source is fed to the planner as a goal;

the planner constructs a plan fragment as a potential corresponding to at least two current consecutive pages;

the planner generates a current optimization criterion from the plan fragment and provides the criterion to the control inferencer;

the control inferencer generates coordinate geodesic curves based on the criterion;

the curves are encoded into a direct memory map structure; and

the encoded curves are the compressed data.

14. The agent of claim 8, wherein the agent is used for data decompression of compressed data from a data source.

15. The agent of claim 14, wherein the data source is a video source.

16. The agent of claim 14, wherein the data source is an audio source.

17. The agent of claim 14, wherein

the compressed data is provided to the knowledge model;

the control inferencer uses the data in the knowledge model to construct coordinate geodesics through an integration procedure; and

the geodesics correspond to the uncompressed data.

18. The agent of claim 7, wherein

the planner and the adapter use symbolic pattern matching during computational operation.

19. The agent of claim 7, wherein

the reactive inferencer is a finite state machine and includes means for recursive finite state machine decomposition for computational operation.

20. The agent of claim 19, wherein the means for recursive finite state machine decomposition further comprises:

means for parallel composition for transforming the finite state machine into a parallel series of unitary automatons;

means for cascade decomposition for transforming each unitary automaton into a series of prefix automatons and loop automatons;

means for linearization for modifying the loop automatons by including a new terminal state to form base automatons;

means for trimming inaccessible base automatons that include inaccessible states;

means for determining if each base automaton corresponds to a successful path in the finite state machine; and

means for outputting each successful base automaton as to the adapter for eventual use as correction terms for the plan.

21. The agent of claim 7, wherein

the adapter receives the failure terms from the control inferencer and computes correction terms as a function of the failure terms;

the planner receives the correction terms from the adapter and generates a new plan from the correction terms; and

the reactive inferencer receives and tests the new plan for consistency with ELL rules stored with the reactive inferencer, if the new plan is inconsistent, then the new plan is sent back to the adapter for refinement, if the new plan is consistent then the new plan is sent to the control inferencer.

22. The agent of claim 7, wherein

the knowledge model uses memory patches to store the requirements of the local process.

23. The agent of claim 1, wherein the control loop is a direct memory map machine.

24. The agent of claim 23, wherein the direct memory map machine comprises:

an input for providing input information from sensory information about the process being controlled and constraint information from other agents;

an array of memory patches for storing a knowledge model corresponding to the input information;

a goal comparand register for comparing a stored process goal with the input information, the register operates to activate at least one of the memory patches of the array of memory patches if the pattern of the process goal at least partially matches the pattern of said at least one of the array of memory patches;

a present state comparand register for comparing a stored present state of the process with the input information, the register operates to activates at least one of the memory patches of the array of memory patches if the pattern of the present state at least partially matches the pattern of said at least one of the array of memory patches;

a respondents buffer for storing matched memory patch contents that are activated by the present state comparand register, the contents represent the current state of the process;

a back propagator buffer for storing matched memory patch contents that are activated by the goal comparand register;

means for unifying a portion of the information stored in the respondents buffer to form a portion of a respondent memory patch;

means for computing the remaining portion of the respondent memory patch using a unique chattering combination of the contents of the respondents buffer and the contents of the back propagator buffer; and

means, using the respondent memory patch, to generate a required control action for controlling the process.

25. The agent of claim 24, wherein the array of memory patches operates as content addressable memory.

26. The agent of claim 24, wherein the memory patch is a computational representation of an open set in a topology of a carrier manifold.