GPGP and TMS Jos e M Vidal Department of Computer Science and - PowerPoint PPT Presentation

GPGP and TÆMS GPGP and TÆMS Jos´ e M Vidal Department of Computer Science and Engineering, University of South Carolina November 7, 2007 Abstract Chapter 9

GPGP and TÆMS Introduction Overview TÆMS: language for representing task hierarchy. GPGP: scheduling and negotiation algorithm. Used by Lesser and students for decades on different problems. Victor Lesser, U. Mass. Amherst.

GPGP and TÆMS TÆMS G 0 and G 1 G 2 G 3 or and G 21 G 22 G 23 G 31

GPGP and TÆMS TÆMS G 0 and G 1 G 2 G 3 or and G 21 G 22 G 23 G 31 resource 1 resource 2 data 1 data 2

GPGP and TÆMS TÆMS G 0 and enables G 1 G 2 G 3 enables or and G 21 G 22 G 23 G 31 resource 1 resource 2 data 1 data 2

GPGP and TÆMS TÆMS G 0 quality : (.2,0)(.8,8) cost : (1,0) duration : (1,2) and enables G 1 G 2 G 3 enables or and G 21 G 22 G 23 G 31 q : (.1,0)(.9,5) c : (1,10) d : (.4,2)(.6,5) resource 1 resource 2 data 1 data 2

GPGP and TÆMS TÆMS Quality Accumulation Functions q min minimum quality of all subtasks q max maximum quality of all subtasks q sum aggregate quality of all subtasks q last quality of most recently completed subtask q sum all as with q sum but all subtasks must be completed q seq min as with q min but all subtasks must be completed in order q seq max as with q max but all subtasks must be completed in order

GPGP and TÆMS GPGP Agent Architecture Design-to-Criteria Schedule Execution Scheduler TÆMS Structure Task and Assessor Goal Criteria Non-Local Commitment Database GPGP Coordination

GPGP and TÆMS GPGP Agent Architecture Design-to-Criteria Schedule Execution Scheduler U s e s TÆMS Structure Task and Assessor Goal Criteria Non-Local Commitment Database GPGP Coordination

GPGP and TÆMS GPGP Agent Architecture Design-to-Criteria Schedule Execution Scheduler U s e s Uses TÆMS Structure Task and Assessor Goal Criteria Non-Local Commitment Database GPGP Coordination

GPGP and TÆMS GPGP Agent Architecture Produces Design-to-Criteria Schedule Execution Scheduler U s e s Uses TÆMS Structure Task and Assessor Goal Criteria Non-Local Commitment Database GPGP Coordination

GPGP and TÆMS GPGP Agent Architecture Produces Uses Design-to-Criteria Schedule Execution Scheduler U s e s Uses TÆMS Structure Task and Assessor Goal Criteria Non-Local Commitment Database GPGP Coordination

GPGP and TÆMS GPGP Agent Architecture Produces Uses Design-to-Criteria Schedule Execution Scheduler U s e s Uses TÆMS Structure Task and Assessor Goal Criteria Action/Sense Non-Local Domain Info. Msgs. Commitment Database GPGP Coordination

GPGP and TÆMS GPGP Agent Architecture Produces Uses Design-to-Criteria Schedule Execution Scheduler U s Updates State Reschedule Requests e s Uses TÆMS Structure Task and Assessor Goal Criteria Action/Sense Non-Local Domain Info. Msgs. Commitment Database GPGP Coordination

GPGP and TÆMS GPGP Agent Architecture Produces Uses Design-to-Criteria Schedule Execution Scheduler U s Updates State Reschedule Requests e s Uses TÆMS Structure Updates Task and Assessor Goal Criteria Action/Sense Non-Local Domain Info. Msgs. Commitment Database GPGP Coordination

GPGP and TÆMS GPGP Agent Architecture Produces Uses Design-to-Criteria Schedule Execution Scheduler U s Updates State Reschedule Requests e s Reschedule Requests Uses TÆMS Structure Updates Task and Assessor Goal Criteria Updates Action/Sense Non-Local Domain Info. Msgs. Commitment Updates Database Coordination Msgs GPGP Coordination

GPGP and TÆMS GPGP Coordination Coordination Relationships G 1 0 G 1 , 2 G 2 2 1

GPGP and TÆMS GPGP Coordination Coordination Relationships G 1 0 G 2 G ∗ 1 2 max G 1 G 2 3 4

GPGP and TÆMS GPGP Coordination Coordination Relationships G 1 G 1 0 0 G 1 G 2 G 2 1 1 2 G 1 G 2 3 4

GPGP and TÆMS GPGP Coordination Coordination Relationships G 1 G 1 0 0 CR G 1 G 2 G 2 1 1 2 G 1 G 2 3 4

GPGP and TÆMS GPGP Coordination Coordination Relationships G 1 G 1 0 0 CR G 1 G 2 G 2 1 1 2 CR CR G 1 G 2 3 4

GPGP and TÆMS GPGP Coordination CR Reasons A non-local effect in the original graph now starts in one graph and ends in another, or a non-local effect or a subtask relationship has one end in one subgraph but the other end in both subgraphs.

GPGP and TÆMS GPGP Coordination Commitment Messages: from 1 to 2 Commit ( Do ( G 1 )) Commit ( Do ( G 3 ))

GPGP and TÆMS GPGP Design-to-Criteria Scheduler Design-to-Criteria Uses search and heuristics. Complex. Needs to create schedules for the coordination module: what if? scenarios. Must find best schedule if many are possible.

GPGP and TÆMS Summary Key Concepts Coordination as distributed optimization—quantitative view of coordination.

GPGP and TÆMS Summary Key Concepts Coordination as distributed optimization—quantitative view of coordination. Family of coordination mechanism for situation-specific control.

GPGP and TÆMS Summary Key Concepts Coordination as distributed optimization—quantitative view of coordination. Family of coordination mechanism for situation-specific control. Domain-independent representation of agent tasks, using TÆMS.

GPGP and TÆMS Summary Key Concepts Coordination as distributed optimization—quantitative view of coordination. Family of coordination mechanism for situation-specific control. Domain-independent representation of agent tasks, using TÆMS. Quantitative coordination relationships among tasks.

GPGP and TÆMS Summary Key Concepts Coordination as distributed optimization—quantitative view of coordination. Family of coordination mechanism for situation-specific control. Domain-independent representation of agent tasks, using TÆMS. Quantitative coordination relationships among tasks. Multiple goals of varying worth. Different deadlines and alternative ways of being solved.

GPGP and TÆMS Summary Key Concepts Coordination as distributed optimization—quantitative view of coordination. Family of coordination mechanism for situation-specific control. Domain-independent representation of agent tasks, using TÆMS. Quantitative coordination relationships among tasks. Multiple goals of varying worth. Different deadlines and alternative ways of being solved. Modular interface between local agent control (planning and scheduling) and coordination mechanisms.