Transcript MASS, JAF and SRTA: Designing a General Agent Environment
SRTA:
The Soft-Real Time Agent Control Architecture
Bryan Horling, Victor Lesser, Regis Vincent, Thomas Wagner presented by Anita Raja
Agent Control
Most multi-agent research addresses inter agent activities The intra-agent mechanics are just as important… Control affects the potential level of flexibility and sophistication for entire agent Generality, efficiency, reliability… Solid general control architecture provides foundation for further research
Motivation
Several existing research artifacts… Task modeling Planning and scheduling Coordination Previous work done in simulation New more demanding domain (ANTs) Real-time Uncertain Resource-bound More realistic conditions Desire to merge these technologies into a cohesive, functional, reusable entity
Soft Real-Time Architecture
Plan and schedule to solve goals Resource constraints Task interaction constraints Deadlines and earliest start times Merge new goals with existing ones Adjust interleaved schedules as necessary Handle unexpected deviations in execution Address time-related failures Resolve conflicts from failed actions
Soft Real-Time
Hard real-time: formally bound and quantitatively describe performance Soft real-time is a looser metric Tasks may still have value if time bounds are exceeded by small amount Our interest is to be statistically “fast enough” Can target more uncertain domains Better handle unexpected events For motivating domain, tasks should be performed within ±500ms of scheduled time
SRTA Context
Operates at the middle agent layer API formed of two parts… Function accessors TÆMS modeling language Comprised of several components Co-exists in JAF framework with other components Domain Problem Solver Soft Real Time Architecture JAF Controller
TÆMS Task Structures
TÆMS is a goal decomposition planning language Tasks represent goals or sub-goals Methods are primitive actions that can be performed QAFs dictate how tasks accrue quality Interrelationships specify interactions between nodes
Java Agent Framework
Component-based design, similar to JavaBeans Individual components are well-encapsulated and potentially ‘autonomous’ Components organized much like a miniature multi-agent system Intra-agent interactions in the form of Direct method invocation Indirect common data handling Event delivery and receipt
Soft Real-Time Architecture
Other Agents Negotiation Problem solver
Goal Description
TÆMS Library
TÆMS Structure Update Expectations
Resource Modeler
Resource Uses
Conflict Resolution Module
Schedule Failure Multiple Structures
Task Merging DTC-Planner Learning
Linear Plan
Partial Order Scheduler
Parallel Schedule Schedule Failure
Parallel Execution Module
Results
Goal Instantiation
Goals are represented using TÆMS May be dynamically created, or read from static files pTÆMS allows for parameterized, template-like structure definitions (spec_method (label Track-Medium) (agent Agent_A) (supertasks Track) (earliest_start_time 500) (deadline 2000) (outcomes (Outcome (density 1.0) (quality_distribution 5.0 0.5 1.0 0.5) (duration_distribution 750.0 1.0) (cost_distribution 0.0 1.0) ) ) ) (spec_commitment (label commitment-1) (type deadline) (from_agent Agent_A) (to_agent Agent_B) (task Track) (earliest_start_time 500) (deadline 2000) )
Planning
(developed by Tom Wagner)
Goal planning by Design-to-Criteria scheduler Select the most appropriate set of end-to-end actions from a structure Considers action and plan duration, quality, cost, interrelationships, constraints Reasons about mandatory and optional requirements, with respect to desired plan criteria Differentiated by reasoning over ‘soft’ conditions Slider Criteria/ Importance Model Raw Goodness 100% Qualit y Cost Durat ion Thresholds/Limits Qualit y Cost 100% Durat ion Qualit y Certainty Cost Durat ion Certainty Thresholds Qualit y Cost Durat ion 100% Meta Raw Goodness T hresholds/ Limit s Cert aint y Cert aint y T hresholds 0% 66% 33% T hreshold Limit $5.75
Limit 50% 50% T hreshold T hreshold T hreshold 80% 25% 25% 25% 25%
Scheduling
DTC was designed as a single-structure scheduler Multiple goal structures must be merged, or assumed independent Merged structures are larger, slower to schedule Goal independence is an impractical condition A more flexible approach is needed
Partially Ordered Scheduling
(developed by Regis Vincent)
Partial ordered scheduler analyses DTC plans Determines task-based precedence constraints Resource modeler detects resource constraints Builds a precedence graph, used for scheduling and rescheduling Key: Leverage DTC’s existing expertise Structure Init Calibrate Set-Parameters Negotiate-Tracking Send-Message Track-Medium Send-Tracking-Info Static Relationship Send-Result 2500 Dynamic Relationship
Resource Modeler
Creates and maintains timeline of expected uses of resources Distribution based: probabilistic start time, duration and quantity consumed or produced Used by scheduler to find and bind appropriate times for methods Used by execution component to monitor resource level expectations
Schedule Merging
STRA natively supports multiple concurrent, interdependent goals PO Scheduler considers prior precedence graphs when scheduling new tasks Conflicts avoided by “shifting” methods based on graph information Avoids monolithic rescheduling …but retains the flexibility to modify prior scheduling results as needed
Execution
Method execution is assumed to be in parallel Constraints (resource, interrelationships, etc.) are validated before method is started Failed constraints require rescheduling PO Scheduler precedence graphs are again used for quick shifting where possible Results are reported to other components and checked for failures
Future Work
Provide an end-to-end model of performance bounds Add anytime character to techniques Meta-level reasoning system to control level of effort and resource expenditure