JPH02281336A - Distributed constraint satisfaction processing system - Google Patents

Abstract

PURPOSE:To eliminate improper selection by asking another agent for confirmation about an added reason at the time when one agent performs inference with a hypothesis to temporarily select a candidate of a solution. CONSTITUTION:An agent 3 consists of two processing modules of a selector 4 and a communication controller 5, a request memory 6 which is referred and changed by these modules, a select set 7, a message queue 8, and a management system ATMS 9. When the selector 4 in each agent selects the solution satisfying a constraint from plural candidates, one agent uses a hypothesis to perform inference. The controller 5 checks contents of the message queue 8 and transfers them to the ATMS 9 and changes the label to add it to the memory 6 and asks another agent 3 for confirmation. The selector 4 uses information to the reply to select a candidate of the solution and excludes improper selection.

Description

[Detailed Description of the Invention] [Industrial Application Field] The present invention relates to solving distributed constraint satisfaction problems, and particularly to distributed cooperative problem solving systems in which a plurality of computer programs that perform calculations independently cooperate to solve problems. This paper relates to a distributed constraint satisfaction processing method for solving distributed constraint satisfaction problems.

[Conventional technology]

A distributed collaborative problem-solving system is a system in which multiple programs (called agents) that perform calculations independently cooperate to solve problems. In a distributed cooperative 1011n solving system, a problem in which each agent selects a solution that satisfies the constraints as a whole from among multiple solution candidates is called a one-distributed constraint light foot problem.

This is a type of distributed constraint satisfaction problem. A distributed constraint satisfaction problem regarding multiple goals can be described as linear.

・A set of agents A= (A+, A-, , ,, A
Set of k goals G = (gl+19.+gn)・Set of plan fragments P = (p +, , , p - goals, set of dependencies between plan fragments Δ However, element d1 of Δ is a + = P t+eP rzΦ9.

This is the format. Here, a is an element of G or P, and +Ptj is an element of P. Φ is one-ofdis
It is called a junction, and P r+eP tie,,
, we show that only the unique plan fragment in
This meaning is an alternative plan PatΦPigΦ00. for establishing a certain goal or plan fragment a1. exists and one of the alternatives holds true.

This means that a goal or a fragment of a plan is established.

・Set of constraints C1 However, element C2 of C is (P,,△P=t△,,,,)
, where P,1 is an element of P. Also, ko is a negation, and △ is an intersection set. This represents a set of plan fragments that cannot hold simultaneously.

The goal and each plan fragment belong to a unique agent. Solving the zero-variance constraint light foot problem means that each agent chooses from the plan fragments belonging to it such that each agent satisfies the constraints and is consistent with the choices of other agents. It is to select a part of it.

There are two 9-order approaches to solving distributed constraint satisfaction problems regarding multiple goals.

Hierarchical solution method: An agent is placed to compile the overall solution, and other agents solve parts of the problem.The coordinating agent integrates the partial solutions to find the overall solution.

Non-hierarchical solution method: Instead of placing an agent in charge of coordinating the overall solution, each agent communicates equally and makes adjustments to reach an overall consensus.

In the hierarchical solution method, each agent makes an inference using a hypothesis, and the inference result based on the hypothesis is sent to the hypothesis-based data integrity management system (Assumption-Based System).
sed Truth Maintenance
System, hereinafter referred to as rATMsJ), is described in the Proceedings of the 2nd National Conference of the Japanese Society for Artificial Intelligence, (198B). use).

ATMS is a system that expresses data whose truth or falsity is uncertain as a hypothesis, and performs inference using the hypothesis. Regarding ATMS.

For example rArtificial Intelligen
ce J Vol, 28 pp.

127-162 (1986) (j,'
de' Mleer: An Assumption-b
asedTMS).

ATMS manages data integrity; a hypothesis is data that can be true or false. A collection of hypotheses is called an environment. The environment creates a lattice based on the inclusion relationships (subset and superset relationships) of a set of hypotheses. For example, FIG. 3 shows a lattice of the environment when four hypotheses A, B, C, and D exist. When data α is derived from a hypothesis (AI, A2. , , , An ) in environment E, data α holds in environment E. For example, in FIG.

If it is derived from A, B, and C that cheetah, that is, a certain animal is a cheetah, then cheetah is established in the environment (A, B, C). ATMS associates all data with the set of environments in which that data exists. This collection of environments is called a label.

If a conflict occurs in environment E, ATMS nogoo E.
d and remove it from the data label, environment E
Similarly, all environments that are a superset of nogo
od is 0. For example, in FIG. 3, if a contradiction occurs in the environment (C, D), the environment (C.

D) and the environment (A, which is a superset of (C, Di)
, C, D), (B, C, D), (A, B
.

C, D) are marked as x in the diagram (nogood).

When some data is derived from other data.

ATMS calculates the label of new data from the label of the underlying data and associates it with the new data. The environment in which the new data holds is calculated by taking the union of the environments in which the prerequisite data hold. If we assume that cheetah is a cheetah because of its brown color and the presence of has-dark-spots, then the environment in which cheetah is valid is the union of the environments in which the prerequisite data is valid.

That is, the environment is calculated as (A, B, C).

In a distributed cooperative problem solving model using ATMS, multiple agents each have an ATMS independently and exchange the following information. These pieces of information affect the ATMS environment managed by each agent, as conceptually shown in FIG.

Note that FIG. 5 shows an example of communication between agents.

Inference result and its label: The inference result of each agent is
Labeled and sent to other agents. The inference results are sent.

When the environment in which the inference result holds is unknown to the receiving agent.

An appropriate new environment is created. In Figure 5.

A case where an inference result that a certain animal is an un-gulate (ungulate) is sent, and the receiving agent does not control the environment [Hl, H4) in which the inference result holds.

New environment [1(,,)1. ) is created.

nogood: If nogood is sent from another agent, < nogood
If (H3) is sent, this nogood supercent environment, i.e. (Hl.

H31, (H2, H31, (Hl, Hz, Hs l is n
It becomes ogood.

On the other hand, without having a coordinating agent.

Multi-stage
The negotiation is 'ReadingsDistri.
Butted Artificial Intellig
encePublished by J Morgan Kaufman, ^lan
11. Bond and Les Ga55eril
(1988) (Conry+S,
E., Meyer.

R, E, and Le3ser+ V, R,:
Multistage Negotiation i
n Distributed Planning),
Multi-stage negotiation is a method for solving a distributed constraint satisfaction problem based on one local knowledge with a small amount of communication, and is described as first-order.

☆1. The agent to which the goal belongs starts the action,
Make a temporary choice on one of the pieces of the plan that leads to the goal.

☆2. Agents ask other agents to confirm that their plan fragment selections are consistent with other agents.☆3. Each agent examines the message queue. Messages are confirmation requests and responses to confirmation requests.

☆４． Each agent examines the confirmation request and the response to the confirmation request, and makes a new temporary selection based on these.

☆5. - If the temporal selection is changed, send a message reflecting the change to the relevant agent. If the confirmation request cannot be satisfied, specify the constraint that caused it to be impossible.

Communicate with the agent who sent the confirmation request.

☆6. Repeat steps ☆3 to ☆5 until the process is completed.

[Problem that the invention is trying to solve]

Hierarchical solution methods require the coordinating agent to perform a lot of processing, and it is expected that the coordinating agent will become a bottleneck in the processing.

On the other hand, in multi-stage negotiation, which is a non-hierarchical solution method, each agent is not given the information necessary to make a choice, so each agent makes many useless choices through trial and error. is expected.

An object of the present invention is to correct these problems and provide a non-hierarchical solution to provide information to each agent to eliminate useless choices, thereby providing a method for avoiding inappropriate choices.

[Means and operations for solving the problem] The present invention performs inference using a hypothesis when finding a solution using multi-stage negotiation, and each agent performs inference on a fragment of the plan.

The feature is that by transmitting the reasoning that a fragment of the plan is required and having other agents refer to the label and make a selection, an inappropriate selection can be avoided and a solution can be efficiently obtained. It is.

〔Example〕

An embodiment of the present invention will be described below with reference to the drawings.

FIG. 1 shows the hardware configuration for realizing a distributed collaborative problem solving system. A plurality of processors l having independent memories 2 are connected by a network and can send messages to each other. The agent is implemented as a program on each processor.

Agent 3 shows an example of the agent configuration in Figure 2.
9, two processing modules, a selector 4 and a communication controller 5, a request memory 6, which is a memory that these modules refer to and change, a selection set 7, a message queue 8, and a
Consists of TMS9.

Figure 6 shows an example of a distributed constraint satisfaction problem regarding multiple goals, and below we will explain the processing flow when there are plan fragments and agents as shown in Figure 6. express the problem.

- Generate hypotheses corresponding to the goals and alternatives to the plan fragments held by each agent. These are called the goal hypothesis and the alternative hypothesis, respectively.

・Express dependencies and constraints of plan fragments using justification.

FIG. 7 shows a set of hypotheses and justifications corresponding to FIG. 6 generated by this method.

For example, as shown in Figure 6, Agene A recognizes that there are two goals and that goal-1 has two alternatives. Agent A has a hypothesis H□+HG corresponding to each goal! and the hypothesis corresponding to the alternative to achieve the goal+(IA, I(za), the label of the fragment IA of the plan is ((Hcl, Hta)) by Justification.

The label for 2A is ((Hat, HA, )), and the label for 3A is ((Hat) ).

Declare choose (HIA, HtA). The meaning of this declaration is to inform ATMS that either hypothesis H+a or HtA is definitely true.

The agent performs the following processing, which is described in connection with multi-stage negotiation☆3.
Corresponds to or ☆6. That is.

☆11. The communication controller examines the contents of the message queue. The contents of the message queue are confirmation requests, cancellation of confirmation requests, and replies to confirmation requests (OK + nogoo).
d). The following processing is performed on the contents of the message queue depending on the contents.

Confirmation request: Pass to ATMS, change label, add to request memory.

Reply to confirmation request (nogood): AT
Add to MS. Add information that the confirmation request was rejected to the confirmation request in the request memory.

Response to request for approval (OK): O in request memory
Add information that it is K.

Cancel confirmation request: Delete the confirmation request from the request memory.

☆120 selector calculates a new selection set.

☆13. A communications controller sends messages to other agents in response to the new selection set. If a confirmation request sent from another agent is rejected, send a corresponding nogood. If the confirmation request sent from another agent is satisfied, send an OK request to the other agent in response to the 0 selection set,
Label and send plan fragments.

☆14. Repeat steps ☆11 to ☆13 until the process is completed.

☆When calculating the new selection set shown in section 12.

The following standards will be established.

The selection criteria is a superset of the environment in which all plan fragments in the two-choice set hold true, contains all goal hypotheses recognized by the agent, and is consistent (nogoo
d) environment must exist.

This criterion means that one should not select plan fragments that contradict the goal hypothesis.

The progress of communication between agents A and B is shown in FIG. 8, and the progress of one communication will be explained below.

5tepl: Agent A chooses 2A from the two alternatives to achieve goal-1 and makes a temporary choice for 2A and 3A. Fragments of these two plans are
It is established in the environment (HGI, HGzHzA). This environment contains all the goal hypotheses and therefore satisfies the criteria mentioned above. A confirmation request is sent to AGEN 1-B.

3tep2: In agent B, 2A and 3A request that IB and 2B be achieved, respectively. IB
, 2B cannot be satisfied simultaneously due to constraints, so
nogood (H(, +, Hcz.

H□) is obtained.

Since the label of 1B is + (HGI, Hzil l, Hc +, I (G!) and all environments in which IB is true are nogood, so the standard cannot be met. This satisfies IB. This means that it is impossible to satisfy both goal-Lgoal-2.Agent B
selects the environment (Hc+, HGd).

Step3: Agent A uses ATMS to obtain nogood (Hc+, HGt,
HzA) and choose (HIA+ HzAl).

Change the label of IA from ((HGl, HlAl) to 1(
G1゜l+ ) (G!), [HGl, H,
A) Change to ). This means that either HIA or Hta must be true, and Hc, 1.
Hat.

If the combination of HIA is inconsistent, HGlr
This reflects the fact that if l(at holds true, then HIA holds true, and as a result, IA is goal-
It is expressed that it is essential to achieve both goals 1 and 2.

Agent A selects IA and 3A.

These two plan fragments are defined by the environment (HGl, Hc, t
) holds, this environment includes all goal hypotheses, and agent A issues a confirmation request to agent C.

The advantages of the present invention can be summarized as follows.

- The meaning of communication between agents has been clarified.

Inappropriate resource allocation can be prevented. ATMS
If agent B is not used, there is no clear criterion for making a decision between IB and 2B, so there is a possibility that agent B will choose IB, and if agent B chooses IB, goal- By using ATMS, it becomes impossible to achieve 2B.
It is communicated that it is essential for the achievement of oal-2, eliminating the possibility of selecting IB.

Also, when not using ATMS, agent B
Even if B was selected, this is the reason why agent A did not select IB.

Simply sending 2AΔ2B −+conflict will cause A to send 2A to goa.
It cannot be concluded that IA is incompatible with 1-1 or that IA is indispensable to achieve the two goals. In the present invention, the fact that IA is essential for achieving two goals is expressed by changing the label.

〔Effect of the invention〕

As explained above, according to the present invention, when each agent performs inference using a hypothesis and each agent temporarily selects a solution candidate, the reason why the solution candidate is needed Inappropriate selections can be eliminated by adding a tag and requesting confirmation from other agents.

[Brief explanation of drawings]

Figure 1 shows an example of the hardware configuration of a distributed collaborative problem solving system, Figure 2 shows an example of the agent configuration, Figure 3 shows an example of the environment lattice, and Figure 4 shows the new A diagram showing an example of the label of the derived data and the label of the data that is the premise for deriving the data. Figure 5 is a diagram showing an example of communication between agents in a distributed collaborative problem solving system. Figure 6 is a diagram showing an example of communication between agents in a distributed collaborative problem solving system. FIG. 7 is a diagram showing an example of a distributed constraint satisfaction problem, FIG. 7 is a diagram expressing the example problem of FIG. 6 using a hypothesis and justification, and FIG. 8 is a diagram showing an outline of communication between agents. In the figure, 1 is a processor, 2 is a memory, 3 is an agent, 4 is a selector, 5 is a communication controller, 6 is a request memory, 7 is a selection set, 8 is a message key, and 9 is ATMS
represents. Patent applicant Nippon Telegraph and Telephone Corporation

Claims (1)

[Claims] In a distributed collaborative problem solving system in which multiple agents, which are computer programs that perform calculations independently, cooperate to solve a problem, each agent selects from multiple solution candidates to satisfy the overall constraint. Distributed constraint satisfaction is characterized in that one agent performs inference using a hypothesis and uses information from other agents to select solution candidates that do not cause constraint violations. Processing method.