1 / 27

Decentralised Structural Reorganisation in Agent Organisations

Ramachandra Kota. Decentralised Structural Reorganisation in Agent Organisations. Motivation. Autonomic systems computing systems with self-management solution to the problem of maintaining large, complex computing systems? (Kephart and Chess, 2003) ‏ Self-organising multi-agent systems

Download Presentation

Decentralised Structural Reorganisation in Agent Organisations

An Image/Link below is provided (as is) to download presentation Download Policy: Content on the Website is provided to you AS IS for your information and personal use and may not be sold / licensed / shared on other websites without getting consent from its author. Content is provided to you AS IS for your information and personal use only. Download presentation by click this link. While downloading, if for some reason you are not able to download a presentation, the publisher may have deleted the file from their server. During download, if you can't get a presentation, the file might be deleted by the publisher.

E N D

Presentation Transcript


  1. Ramachandra Kota Decentralised Structural Reorganisation in Agent Organisations

  2. Motivation • Autonomic systems • computing systems with self-management • solution to the problem of maintaining large, complex computing systems? (Kephart and Chess, 2003)‏ • Self-organising multi-agent systems • autonomous, adaptive and robust • a paradigm to develop autonomic systems (Tesauro et al., 2004)‏

  3. Self-Organisation: Characteristics(Di Marzo Serugendo et al., 2005, 2006)‏ • No External Control – autonomous • Dynamic Operation – continuous over time • No Central Authority – decentralised and robust

  4. Problem Solving Agent Organisations • We need agent systems which can be mapped onto computing systems that perform tasks • We focus on multi-agent systems that act as a problem solving organisation • organisations that receive inputs, perform tasks and return results

  5. Research Objective “Develop a decentralised reorganisation method that can be employed by the agents in a problem solving agent organisation to improve the performance of the organisation as a whole.” • can be used by any agent at any level of the organisation, at any time. • focus on changing the organisational characteristics rather than the agents themselves

  6. Self-organisation approaches • Stigmergic • self-organisation emerges through indirect interactions of the agents (Mano et al., 2006)‏ • Organisational Self Design (OSD)‏ • splitting and merging of agents to achieve reorganisation Gasser and Ishida (1991), Kamboj and Decker (2006) • Adaptive Multi-Agent Systems theory (AMAS)‏ • agents perceive non-cooperative situations (pre-specified) and take rectifying measures. (Capera et al., 2003)

  7. Other Reorganisation Approaches • Diagnostic Subsystem in Agents (Horling et al. 2001)‏ • a diagnostic system that detects the need for reorganisation • MOISE+ controlled reorganisation (Hubner et al. 2004)‏ • a top-down approach using specialised agents • Max-flow network approach (Hoogendoorn 2007)‏ • a centralised solution to resolve bottle-necks There is no existing decentralised mechanism to improve the performance of an organisation composed of invariant agents.

  8. Agent Organisation Model • To act as a framework on which to base our reorganisation method • Existing models: • Moise, Islander, VDT, Opera, Omni etc • We pick up ideas from several models to develop a simple framework

  9. Our Model: Agents • Problem solving agents • receive a task • assign its dependencies and obtain the results of their execution • execute the task and return the result. • Invariant and cooperative agents • Provide a set of services (SA)‏ • Have limited computational capacity (LA)‏ • Example: • Agent A = < SA , LA > where SA ={a, b}, LA = 10 computational units • Agent B = < SB , LB > where SB ={b, c, d} LB = 15 computational units

  10. Our Model: Tasks S0 [a, 4, 5] • Tree structure • Every node represents a service instance • A service instance specifies • type of service • computational units per time-step • number of time-steps required • Dependency - a node can be executed only after the completion of all its child-nodes S1 [b, 3, 9] S2 [c, 5, 2] S3 [a, 8, 6] S4 [d, 2, 3]

  11. Our Model: Organisation Structure • Structure is based on the relationships between the agents • Relation between two agents determines the kind of interaction possible between them • Three kinds of relationships:- • Acquaintance: no interaction • Peer: weak interaction • Authority (superior-subordinate): strong interaction

  12. Our Model: Agent Relations • All agents are acquaintances of each other • Accumulated Service Set: the union of the service set of the agent and the service sets of its subordinates. • Agents are aware of • the personal service sets of their peers • the accumulated service sets of their subordinates X Z Y W

  13. Organisation at work: an example S0 [a, 4, 5] X S1 [b, 3, 9] S2 [c, 5, 2] Z Y S3 [a, 8, 6] S4 [d, 2, 3] W Task Organisation

  14. Evaluation Mechanism 1/3 Agents have to perform two kinds of actions • Allocation of service instances (management) • Execution of service instances • Load on agent x: lx = ∑ (rix + M.mix) • rix is the amount of processing computation of x required by task Ti, • mix is the amount of management computation done by x for task Ti • TxE is the set of tasks being executed by x • M is the management load coefficient • lx <= Lx ; excess tasks will be in the waiting queue TxW |TxE| i=0

  15. Evaluation Mechanism 2/3 • Performance is determined by cost and benefit of the organisation, calculated at every time step. • Cost of agent x: Costx = Lx + C.cx • Lx is capacity of agent x • cx is the number of messages sent by x • C is communication cost coefficient • Cost of the Organisation: Costorg = C. ∑cx + ∑ Lx • A is the set of agents A A x=0 x=0

  16. Evaluation Mechanism 3/3 |TxE| |TxW| • Benefit from x: Benefitx = ∑rix - ∑rix • rix is the amount of computation required by task Ti being executed by x • TxE is the set of tasks being executed by x • TxW is the set of tasks waiting to be executed by x • Benefit of the Organisation: Benefitorg = ∑ Benefitx i=0 i=0 |A| x=0

  17. Reorganisation - scenario S0 [a, 4, 5] X X S1 [b, 3, 9] S2 [c, 5, 2] Z Z Y Y S3 [a, 8, 6] S4 [d, 2, 3] W W Task Organisation

  18. Reorganisation Method: Actions • Formulated using the decision theoretic approach • Changing the relation – denoted as actions Peers Subordinate Subordinate Subordinate Peers Just acquaintances Just acquaintances

  19. Reorganisation Method: Value function • Pairs of agents jointly estimate the expected utility of changing their relation • A combined Value function of the form: Vx,y = ΔLoadx+ΔLoady+ΔLoadOA+ΔCostcomm+Costreorg • Value is calculated for every possible action in the state and the action with maximum expected value is chosen.

  20. Attribute values: FORM_SUBR(x,y) action • ΔLoadx = - Asgx,Tot * M * filledx(ttotal) / ttotal • ΔLoady = - AsgLOAD * M * filledy(tsubr) * ttotal / (tsubr)2 • ΔLoadOA = OALOAD [load on other agents] • ΔCostcomm= OACOST [cost because of other agents] • Costreorg = - R [reorganisation cost constant] x,y x,y x,y x,y x,y The attribute values are calculated on basis of past interactions and delegations involving the two agents

  21. Experimental Evaluation • Compare our method with a random reorganisation strategy. • Random strategy: An agent randomly chooses to change some of its relations • Performance is evaluated on basis of the average cost and benefit obtained from the simulation runs

  22. Simulation Parameters 1/2 • Distribution of Services: • agents may have distinct service sets or overlapping service sets • determined by ‘service probability’ (sp) • sp = 0 : every agent has a unique service set • sp = 1 : every agent can perform all services

  23. Simulation Parameters 1/2 • Similarity between Tasks: • could be completely unrelated • could be composed of a finite set of constituents (Patterns)

  24. Results 1/2 Dissimilar Tasks Similar Tasks

  25. Results 2/2 Distinct service sets Highly overlapping service sets

  26. Future Work • Upper bound: • an oracle organisation with complete information of the future tasks • a centralised reorganiser/allocator • Efficient Reorganisation • compute utilities for a selective set of relations only, at a given time • Dynamic agents, organisation norms etc.

  27. Thank you!! ??

More Related