Agent Based Production Planning

1
Agent Based Production Planning
intro agent-based production planning decompos
ition techniquesSBC/ISBC ExPlanTech
MAS CPlanT MAS conclusions

• Michal Pechoucek
• Gerstner Laboratory, Czech Technical University

2
Agent Based Systems
intro

• agent is an encapsulated computational system,
  that is situated in some environment, and that is
  capable of flexible, autonomous behaviour in
  order to meet its design objective (Wooldridge).
• an agent is not only an object, process, program,
  situated robot, ..
• critical difference agents internal decision
  making processes are not transparent one cannot
  prove what the other agent will do.
• this property (and fact that agents are usually
  developed by different developers) causes
  emergent behaviour that has not been thought of
  at the design time
• agents can be standalone or members of a
  multi-agent system
• distributed artificial intelligence is a branch
  of science that studies social aspects of
  artificial intelligence, e.g. communication,
  cooperation, collective mental states
• multi-agent system is a collection of agents that
  work together in order to meet an
  in-community-shared goal
• agent based system is a system whose
  functionality is based on operation of agent(s),
  which may be of collaborative or self-interested
  nature

3
What can Agents Provide Production Planning with?

• design architecture (e.g. Prosa architecture,
  Holonic Manufacturing Systems, ProPlanT
  architecture, etc.)
• integration/agentification technology (e.g. FIPA
  standards, agent development environments)
• planning algorithms distributed decision making
  (e.g. stigmergy, negotiation and auctioning,
  social intelligence based interaction, etc.)

4
Agent-based Production Planning
agent-based production planning

• advantages of agent-based planning approaches
• reconfigurability and flexibility, tractability
  (distributed), higher degree of planning
  efficiency
• there are three fundamental approaches to
  agent-based planning
• decomposition based planning there is a
  temporary or permanent hierarchy of agents where
  each decomposes a task into subtasks and
  coordinates its completion. can be done via
  contract-net-protocols, subscriptions, etc.
• fully autonomous planning all agents see the
  planning problem and form their local plans.
  these plans are later merged and conflicts are
  resolved by re-planning e.g. PGP Partial
  Global Planning. agents share a common knowledge
  structure (blackboard) or there is a high-level
  coordinator (who resolves the conflicts) or
  agents interact via rather inefficient
  distributed techniques (negotiation, broadcast,
  rings, voting, etc.)
• backward chaining planning a compromise
  between (i) and (ii). the request backpropagets
  in the manufacturing flow. there is no
  command-and-control hierarchy and no central
  component, but agents negotiate via
  contract-net-protocols, subscriptions, etc.

5
Decomposition Based Planning

• we want to arrive at a distributed plan that will
  achieve a high-level task
• each task ? can be planned either by means of a
• team action plan result of inter-agent
  negotiation and mutual agreeing upon joint
  commitments or
• individual plan shall implement a single
  agents commitment (planning by linear/non-linear
  planning)
• the problem is to decide
• how to decompose a task into subtask
• whom to subcontract for cooperation

6
Team Action Plan

• team action plan ?(?) is as a set ?(?) ??i,
  Aj, start(?i), due(?i), price(?i)?.
• ?(?) is correct if all the collaborators Aj are
  able to implement the task ?j in the given time
  and for the given price.
• ?(?) is accepted if all agents Aj get committed
  to implementing the task ?j in the given time and
  for the given price.
• ? is achievable, if there exists such ?(?) that
  is correct.
• ? is planned, if there exists ?(?) that is
  accepted

7
Individual Action Plan

• individual plan ?(?) is as either an unordered
  set
• ?(?) ??i, start(?i), due(?i), price(?i)?.
• or a partially ordered set
• ?(?) ??i, price(?i)?.
• ?(?) is correct (complete and consistent) if it
  is executable and implements ?.
• ?(?) is complete iff all the preconditions of the
  operators are satisfied by an effect of another
  operator (or by initial conditions).
• ?(?) plan is consistent iff ordering among
  operators does not contradict or operators from
  the same world do not provide contradicting
  effects

8
Decomposition/Contraction Techniques
decomposition techniquesSBC/ISBC

• contract-net-protocol (CNP)
• auctions
• subscription based contraction (SBC)
• iterated SBC (I-SBC)

9
Decomposition/Contraction Techniques

• contract-net-protocol (CNP)
• auctions
• subscription based contraction (SBC)
• iterated SBC (I-SBC)
• auction protocols
• English (first-price open-cry) sometimes an
  open-exit
• sealed-bid first-price
• Dutch auction
• Vickery (sealed-bid second-price)
• all-pay auctions (computer science)

10
Decomposition/Contraction Techniques

• subscription based contraction (SBC)

11
Social Knowledge (SK)

• agents knowledge is either
• problem solving knowledge asocial type of
  skill guide agents autonomous local decision
  making processes (aimed e.g. at providing an
  expertise or search in the agents database)
• self knowledge knowledge about agents
  behavior, status and commitments (a special
  instance of social knowledge below)
• social knowledge knowledge about other agents,
  their behavioral patterns, their capabilities,
  load, experiences, commitments, but also
  knowledge and belief
• social knowledge is located in agents wrapper
  in an acquaintance model

communication layer
wrapper
acquaintance model
body
body
12
Tri-base Acquaintance Model

• acquaintance model is a computational model of
  agents mutual awareness, it stores and
  maintains agents social knowledge
• decomposition on request
• exploitation of the pre-prepared plan
• new plan generation (based on SB knowledge)
• new plan generation (broadcasting)
• replanning driven by state-base update

13
CF Acquaintance Model

• Soc-BB(A0)KS(Ai) for ?Ai? ?(A0),
  Com-BB(A0)Kp(Ai) for ?Ai ? ?(A0)
• Self-BB(A0) Kp(A0), KS(A0), KPr(A0),
  Coal-BB(A0) coalitions, rules
• reduces the communication traffic and thus the
  increases problem solving efficiency, while it
  requires substantial communication for the
  acquaintance model maintenance

14
Example
15
Decomposition/Contraction Techniques

• contract-net-protocol (CNP)
• auctions
• subscription based contraction (SBC)
• iterated SBC (I-SBC)
• SBC difficulties
• maintenance too much of data, how often,
• monitoring selectivity
• frequency of requests
• still high complexity on the side of the
  coordinator

16
Decomposition/Contraction Techniques

• contract-net-protocol (CNP)
• auctions
• subscription based contraction (SBC)
• iterated SBC (I-SBC)
• therefore we suggest an improvement of SBC that
  is good for very complex domains, where not all
  data are available (confidentiality reasons) or
  there are too much of data (complexity problems)
• exploitation of the concept of the private,
  public and semi-private knowledge (as much as the
  concept alliances), where only approximation of
  the planning data is made available to agents
  social models

17
Iterated SBC (I-SBC)
coordinator
18
Iterated SBC (I-SBC)
coordinator
19
Iterated SBC (I-SBC)
coordinator
20
Iterated SBC (I-SBC)
coordinator
21
Iterated SBC (I-SBC)
agent1
resources
agent2
agent3
t
t
22
Iterated SBC (I-SBC)
agent1
resources
agent2
agent3
t
t
23
Agent-Based Planning in the Gerstner Laboratory
ExPlanTech MAS

• ExPlanTech Production Planning Multi-agent
  System
• CPlanT Coalition Planning Multi-Agent System
  for OOTW planning

24
ExPlanTech Domain Specification

• ExPlanTech a production planning system with a
  functionality to
• estimating due dates and resources requirements
• providing a project plan
• implementing re-planning
• extra-enterprise extension
• to allow remote access
• integrate supply-chain relations

25
ExPlanTech Architecture
26
ExPlanTech Implementation

• operator an instance of the ppa and pma classes
  project configuration and decomposition,
  management of the overall project
• workshop an instance of the pa class
  scheduling and resource allocation on a
  department or CNC machine
• database agent an instance of the pa class an
  integration agent, integrates ExPlanTech with
  factory ERP
• material agent an instance of the pa class
  integrates an MRP - material resource planning
  system
• FIPA compliant system, implemented in JADE (Java
  Agent Development Environment).
• Distributed over several machines, each agent
  has got a GUI for user interaction
• new agents can login and the confuigu-ration can
  be altered in runtime
• Integrated with MS-Project, JDBC, IE
• Special visualization and user manipulation
  meta-agent

27
ExPlanTech ExtraPlanT Exetnsion
28
Agent-Based Planning in the Gerstner Laboratory
CPlanT MAS

• ExPlanTech Production Planning Multi-agent
  System
• CPlanT Coalition Planning Multi-Agent System
  for OOTW planning

29
CPlanT Domain Specification

• domain Operations other than war (OOTW)
  humanitarian relief operations, peace-keeping
  missions, non-combat operations
• each entity/actor (governmental institutions,
  troops, humanitarian bodies, NGOs, charities)
  represented by an agent
• domain specifics (simplified)
• equality anyone can initiate forming a
  coalition no hierarchy
• reluctance to share vital planning information
• agents inaccessibility poor communication
  links,
• collaborative/self interested different
  cultural backgrounds
• key problems
• minimize required communication traffic
  (affects problem solving efficiency)
• keep the quality of the operation the
  coalitions perform reasonably good
• minimize loss of agents private knowledge
  disclosure,
• minimize the amount of the shared information

30
CPlanT Key Ideas

• organizing the agents into alliances (structural
  decomposition)
• a particular task (a mission) accomplished by a
  coalition (preferably created as a subset of an
  alliance)
• structuring the agents private, semi-private,
  public knowledge
• using the concept of the tri-base acquaintance
  model and social intelligence
• designing advanced methods for inter-agent
  negotiation

31
CPlanT Coalition Formation Operation Lifecycle

• Registration central registration of the public
  knowledge
• Alliance Formation communicated via selective
  single-stage CNP
• Coalition Leader Selection collective decision
  making
• Coalition Formation communicated via
  acquaintance models based contraction
• Team Action Planning collective planning of a
  team action combination of CNP and AM

32
CPlanT Implementation
33
Conclusions
conclusions

• agents in production and resource allocation
  planning are good as
• the planning system is scalable and easy to be
  reconfigured
• problem solving efficiency can be increased by an
  appropriate structuring of the community and
  acquaintance model design
• they are efficient in areas with natural
  distribution (e.g. supply chains)
• for handling imprecise information and inexact
  knowledge
• http//agents.felk.cvut.cz
• http//gerstner.felk.cvut.cz