Network-Aware Automated Planning and Plan Execution Kyle Usbeck A - PowerPoint PPT Presentation

Introduction Motivation Formalization Background Technical Approach Approach Experiments Contributions Qualitatively-different plans: 1 Generating plans over a range of evaluation criteria; Visualizing plan evaluations. Improve plan selection. Network-Aware Agents: 2 Classical planning domains for distributed service composition; Measuring the performance and effectiveness of planning, execution, and monitoring agents; Incorporating network-awareness. Department of Science Computer Kyle Usbeck Network-Aware Automated Planning 7/75

Introduction Motivation Formalization Background Technical Approach Approach Experiments Outline 1 Introduction Motivation Background Approach 2 Formalization Problem Statement 3 Technical Approach Planning Agents Execution Agents Monitoring Agents Mixed-initiative UI 4 Experiments Plan Evaluation Benchmarking Network-Aware Agent Combinations Discussion Department of Science Computer Kyle Usbeck Network-Aware Automated Planning 8/75

Introduction Motivation Formalization Background Technical Approach Approach Experiments Service Composition to Automated Planning Definition “Service composition is the linking. . . of existing services so that their aggregate behavior is that of a desired service (the goal)” [Hoffmann et al . 09]. Requires Semantic Web Services [Sirin et al . 04]. QoS Assurance [Gu et al . 03]. Assumes static networking. Department of Science Computer Kyle Usbeck Network-Aware Automated Planning 9/75

Introduction Motivation Formalization Background Technical Approach Approach Experiments Service Composition to Automated Planning Definition “Service composition is the linking. . . of existing services so that their aggregate behavior is that of a desired service (the goal)” [Hoffmann et al . 09]. Requires Semantic Web Services [Sirin et al . 04]. QoS Assurance [Gu et al . 03]. Assumes static networking. Department of Science Computer Kyle Usbeck Network-Aware Automated Planning 9/75

Introduction Motivation Formalization Background Technical Approach Approach Experiments Service Composition to Automated Planning Definition “Service composition is the linking. . . of existing services so that their aggregate behavior is that of a desired service (the goal)” [Hoffmann et al . 09]. Requires Semantic Web Services [Sirin et al . 04]. QoS Assurance [Gu et al . 03]. Assumes static networking. Department of Science Computer Kyle Usbeck Network-Aware Automated Planning 9/75

Introduction Motivation Formalization Background Technical Approach Approach Experiments Agents in Planning Model (Domain) Agents: Current Planner Goal(s) State Planning Agent. Plans Feedback Execution Agent. Controller Sensor Monitoring Agent. [Tate 93] Actions Observations Events System Department of Science Computer Kyle Usbeck Network-Aware Automated Planning 10/75

Introduction Motivation Formalization Background Technical Approach Approach Experiments Agents in Planning Model (Domain) Agents: Current Planner Goal(s) State Planning Agent. Plans Feedback Execution Agent. Controller Sensor Monitoring Agent. [Tate 93] Actions Observations Events System Department of Science Computer Kyle Usbeck Network-Aware Automated Planning 10/75

Introduction Motivation Formalization Background Technical Approach Approach Experiments Agents in Planning Model (Domain) Agents: Current Planner Goal(s) State Planning Agent. Plans Feedback Execution Agent. Controller Sensor Monitoring Agent. [Tate 93] Actions Observations Events System Department of Science Computer Kyle Usbeck Network-Aware Automated Planning 10/75

Introduction Motivation Formalization Background Technical Approach Approach Experiments Planning Under Uncertainty Restrictive Assumptions: Sources of Uncertainty: Determinism. Partial observability. Full observability. Unreliable resources. Reachability goals. Measurement variance. [Nau et al . 04] Inherently vague concepts. Department of Science Computer Kyle Usbeck Network-Aware Automated Planning 11/75

Introduction Motivation Formalization Background Technical Approach Approach Experiments Planning Under Uncertainty Restrictive Assumptions: Sources of Uncertainty: Determinism. Partial observability. Full observability. Unreliable resources. Reachability goals. Measurement variance. [Nau et al . 04] Inherently vague concepts. Department of Science Computer Kyle Usbeck Network-Aware Automated Planning 11/75

Introduction Motivation Formalization Background Technical Approach Approach Experiments Fault Detection & Isolation (FDI) System Types of FDI: Analytic. Residual Data-driven. Generation Knowledge-based. [Pettersson 05] Decision Making Department of Science Computer Kyle Usbeck Network-Aware Automated Planning 12/75

Introduction Motivation Formalization Background Technical Approach Approach Experiments Fault Detection & Isolation (FDI) Types of FDI: Analytic. Data-driven. Knowledge-based. [Pettersson 05] Department of Science Computer Kyle Usbeck Network-Aware Automated Planning 12/75

Introduction Motivation Formalization Background Technical Approach Approach Experiments Fault Detection & Isolation (FDI) Inputs Outputs Types of FDI: Analytic. Data-driven. Knowledge-based. [Pettersson 05] Department of Science Computer Kyle Usbeck Network-Aware Automated Planning 12/75

Introduction Motivation Formalization Background Technical Approach Approach Experiments Outline 1 Introduction Motivation Background Approach 2 Formalization Problem Statement 3 Technical Approach Planning Agents Execution Agents Monitoring Agents Mixed-initiative UI 4 Experiments Plan Evaluation Benchmarking Network-Aware Agent Combinations Discussion Department of Science Computer Kyle Usbeck Network-Aware Automated Planning 13/75

Introduction Motivation Formalization Background Technical Approach Approach Experiments Approach Modify planner to improve the quality of the plans it 1 produces based on evaluation criteria. Add network-awareness to planning, execution, and 2 monitoring agents. Purpose To improve network-centric automated planning and execution. Department of Science Computer Kyle Usbeck Network-Aware Automated Planning 14/75

Introduction Motivation Formalization Background Technical Approach Approach Experiments Approach Modify planner to improve the quality of the plans it 1 produces based on evaluation criteria. Add network-awareness to planning, execution, and 2 monitoring agents. Purpose To improve network-centric automated planning and execution. Department of Science Computer Kyle Usbeck Network-Aware Automated Planning 14/75

Introduction Motivation Formalization Background Technical Approach Approach Experiments Approach Modify planner to improve the quality of the plans it 1 produces based on evaluation criteria. Add network-awareness to planning, execution, and 2 monitoring agents. Purpose To improve network-centric automated planning and execution. Department of Science Computer Kyle Usbeck Network-Aware Automated Planning 14/75

Introduction Formalization Problem Statement Technical Approach Experiments Outline 1 Introduction Motivation Background Approach 2 Formalization Problem Statement 3 Technical Approach Planning Agents Execution Agents Monitoring Agents Mixed-initiative UI 4 Experiments Plan Evaluation Benchmarking Network-Aware Agent Combinations Discussion Department of Science Computer Kyle Usbeck Network-Aware Automated Planning 15/75

Introduction Formalization Problem Statement Technical Approach Experiments Formal Problem Statement Σ is the planning domain — the model of the world passed as input to the planner. Σ is a Tuple S set of states; A set of actions; E set of events; γ transition function γ : S × A → S . Department of Science Computer Kyle Usbeck Network-Aware Automated Planning 16/75

Introduction Formalization Problem Statement Technical Approach Experiments Formal Problem Statement Σ is the planning domain — the model of the world passed as input to the planner. Σ is a Tuple S set of states; A set of actions; E set of events; γ transition function γ : S × A → S . Department of Science Computer Kyle Usbeck Network-Aware Automated Planning 16/75

Introduction Formalization Problem Statement Technical Approach Experiments Formal Problem Statement Σ is the planning domain — the model of the world passed as input to the planner. Σ is a Tuple S set of states; A set of actions; E set of events; γ transition function γ : S × A → S . Department of Science Computer Kyle Usbeck Network-Aware Automated Planning 16/75

Introduction Formalization Problem Statement Technical Approach Experiments Formal Problem Statement Σ is the planning domain — the model of the world passed as input to the planner. Σ is a Tuple S set of states; A set of actions; E set of events; γ transition function γ : S × A → S . Department of Science Computer Kyle Usbeck Network-Aware Automated Planning 16/75

Introduction Formalization Problem Statement Technical Approach Experiments Formal Problem Statement The functions on planning actions: For a ∈ A precond ( a ) preconditions of a ; effects + ( a ) positive effects of a ; effects − ( a ) negative effects of a ; host ( a ) the single host h from a ; resources ( a ) the set of resources (parameters) of action a . Department of Science Computer Kyle Usbeck Network-Aware Automated Planning 17/75

Introduction Formalization Problem Statement Technical Approach Experiments Formal Problem Statement The functions on planning actions: For a ∈ A precond ( a ) preconditions of a ; effects + ( a ) positive effects of a ; effects − ( a ) negative effects of a ; host ( a ) the single host h from a ; resources ( a ) the set of resources (parameters) of action a . Department of Science Computer Kyle Usbeck Network-Aware Automated Planning 17/75

Introduction Formalization Problem Statement Technical Approach Experiments Formal Problem Statement The functions on planning actions: For a ∈ A precond ( a ) preconditions of a ; effects + ( a ) positive effects of a ; effects − ( a ) negative effects of a ; host ( a ) the single host h from a ; resources ( a ) the set of resources (parameters) of action a . Department of Science Computer Kyle Usbeck Network-Aware Automated Planning 17/75

Introduction Formalization Problem Statement Technical Approach Experiments Formal Problem Statement The functions on planning actions: For a ∈ A precond ( a ) preconditions of a ; effects + ( a ) positive effects of a ; effects − ( a ) negative effects of a ; host ( a ) the single host h from a ; resources ( a ) the set of resources (parameters) of action a . Department of Science Computer Kyle Usbeck Network-Aware Automated Planning 17/75

Introduction Formalization Problem Statement Technical Approach Experiments Formal Problem Statement The functions on planning actions: For a ∈ A precond ( a ) preconditions of a ; effects + ( a ) positive effects of a ; effects − ( a ) negative effects of a ; host ( a ) the single host h from a ; resources ( a ) the set of resources (parameters) of action a . Department of Science Computer Kyle Usbeck Network-Aware Automated Planning 17/75

Introduction Formalization Problem Statement Technical Approach Experiments Formal Problem Statement The planning agent receives the tuple, I P , and creates a set of plans, P I . I P is a Tuple Σ automated planning domain; s 0 initial state; S g set of goal state(s); H set of hosts (nodes) on the network; ω H host link weighting. Department of Science Computer Kyle Usbeck Network-Aware Automated Planning 18/75

Introduction Formalization Problem Statement Technical Approach Experiments Formal Problem Statement The planning agent receives the tuple, I P , and creates a set of plans, P I . I P is a Tuple Σ automated planning domain; s 0 initial state; S g set of goal state(s); H set of hosts (nodes) on the network; ω H host link weighting. Department of Science Computer Kyle Usbeck Network-Aware Automated Planning 18/75

Introduction Formalization Problem Statement Technical Approach Experiments Formal Problem Statement The planning agent receives the tuple, I P , and creates a set of plans, P I . I P is a Tuple Σ automated planning domain; s 0 initial state; S g set of goal state(s); H set of hosts (nodes) on the network; ω H host link weighting. Department of Science Computer Kyle Usbeck Network-Aware Automated Planning 18/75

Introduction Formalization Problem Statement Technical Approach Experiments Formal Problem Statement The planning agent receives the tuple, I P , and creates a set of plans, P I . I P is a Tuple Σ automated planning domain; s 0 initial state; S g set of goal state(s); H set of hosts (nodes) on the network; ω H host link weighting. Department of Science Computer Kyle Usbeck Network-Aware Automated Planning 18/75

Introduction Formalization Problem Statement Technical Approach Experiments Formal Problem Statement The planning agent receives the tuple, I P , and creates a set of plans, P I . I P is a Tuple Σ automated planning domain; s 0 initial state; S g set of goal state(s); H set of hosts (nodes) on the network; ω H host link weighting. Department of Science Computer Kyle Usbeck Network-Aware Automated Planning 18/75

Introduction Formalization Problem Statement Technical Approach Experiments Formal Problem Statement Problem To find and execute p I ∈ P I where p I = { a 0 , a 1 , . . . , a | p I | } and execution of p I yields the best domain-dependent and network-centric evaluations. Network-Awareness An agent exhibits network-awareness if changes to ω H cause the agent’s output to change while all other inputs remain constant. Department of Science Computer Kyle Usbeck Network-Aware Automated Planning 19/75

Introduction Formalization Problem Statement Technical Approach Experiments Formal Problem Statement Problem To find and execute p I ∈ P I where p I = { a 0 , a 1 , . . . , a | p I | } and execution of p I yields the best domain-dependent and network-centric evaluations. Network-Awareness An agent exhibits network-awareness if changes to ω H cause the agent’s output to change while all other inputs remain constant. Department of Science Computer Kyle Usbeck Network-Aware Automated Planning 19/75

Introduction Formalization Problem Statement Technical Approach Experiments Formal Problem Statement Host 1 Host 2 Planning Agent Service 1 Execution Service 2 Agent Host 3 Host 4 Service 1 Host 5 Monitoring Agent Monitoring Agent Service 3 Service 2 Service 3 Department of Science Computer Kyle Usbeck Network-Aware Automated Planning 20/75

Introduction Formalization Problem Statement Technical Approach Experiments Formal Problem Statement Host 1 Planning Agent Plans Faults Execution Agent Service Calls Host 2 Host 3 Host 4 Host 5 Department of Science Computer Kyle Usbeck Network-Aware Automated Planning 21/75

Introduction Formalization Problem Statement Technical Approach Experiments Formal Problem Statement Planning Agent Execution Agent Monitoring Agent Plan(s) [minor fault] Fault [major fault] Fault Department of Science Computer Kyle Usbeck Network-Aware Automated Planning 22/75

Introduction Planning Agents Formalization Execution Agents Technical Approach Monitoring Agents Experiments Mixed-initiative UI Outline 1 Introduction Motivation Background Approach 2 Formalization Problem Statement 3 Technical Approach Planning Agents Execution Agents Monitoring Agents Mixed-initiative UI 4 Experiments Plan Evaluation Benchmarking Network-Aware Agent Combinations Discussion Department of Science Computer Kyle Usbeck Network-Aware Automated Planning 23/75

Introduction Planning Agents Formalization Execution Agents Technical Approach Monitoring Agents Experiments Mixed-initiative UI Planning Domain Extensions Operator distribution e.g. , N ODE 1A CTION ( parameters ) Implicit constraints. Resource distribution e.g. , A CTION ( node1 , parameters ) s 0 ← s 0 ∪ { T YPE ( node 1 ) = NETWORK N ODE } s 0 ← s 0 ∪ { A CTION ( node 1 ) = true } Department of Science Computer Kyle Usbeck Network-Aware Automated Planning 24/75

Introduction Planning Agents Formalization Execution Agents Technical Approach Monitoring Agents Experiments Mixed-initiative UI Planning Domain Extensions Operator distribution e.g. , N ODE 1A CTION ( parameters ) Implicit constraints. Resource distribution e.g. , A CTION ( node1 , parameters ) s 0 ← s 0 ∪ { T YPE ( node 1 ) = NETWORK N ODE } s 0 ← s 0 ∪ { A CTION ( node 1 ) = true } Department of Science Computer Kyle Usbeck Network-Aware Automated Planning 24/75

Introduction Planning Agents Formalization Execution Agents Technical Approach Monitoring Agents Experiments Mixed-initiative UI Planning Domain Extensions Operator distribution e.g. , N ODE 1A CTION ( parameters ) Implicit constraints. Resource distribution e.g. , A CTION ( node1 , parameters ) s 0 ← s 0 ∪ { T YPE ( node 1 ) = NETWORK N ODE } s 0 ← s 0 ∪ { A CTION ( node 1 ) = true } Department of Science Computer Kyle Usbeck Network-Aware Automated Planning 24/75

Introduction Planning Agents Formalization Execution Agents Technical Approach Monitoring Agents Experiments Mixed-initiative UI Planning Domain Extensions Operator distribution e.g. , N ODE 1A CTION ( parameters ) Implicit constraints. Resource distribution e.g. , A CTION ( node1 , parameters ) s 0 ← s 0 ∪ { T YPE ( node 1 ) = NETWORK N ODE } s 0 ← s 0 ∪ { A CTION ( node 1 ) = true } Department of Science Computer Kyle Usbeck Network-Aware Automated Planning 24/75

Introduction Planning Agents Formalization Execution Agents Technical Approach Monitoring Agents Experiments Mixed-initiative UI Planning Domain Extensions Operator distribution e.g. , N ODE 1A CTION ( parameters ) Implicit constraints. Resource distribution e.g. , A CTION ( node1 , parameters ) s 0 ← s 0 ∪ { T YPE ( node 1 ) = NETWORK N ODE } s 0 ← s 0 ∪ { A CTION ( node 1 ) = true } Department of Science Computer Kyle Usbeck Network-Aware Automated Planning 24/75

Introduction Planning Agents Formalization Execution Agents Technical Approach Monitoring Agents Experiments Mixed-initiative UI Planning Domain Extensions Operator distribution e.g. , N ODE 1A CTION ( parameters ) Complexity Implicit constraints. Operator distribution increases the number of actions in Σ to | H | × | A | in the Resource distribution worst case. e.g. , A CTION ( node1 , parameters ) Resource distributed increases the s 0 ← s 0 ∪ { T YPE ( node 1 ) = NETWORK N ODE } number of constraints in the world-state. s 0 ← s 0 ∪ { A CTION ( node 1 ) = true } Department of Science Computer Kyle Usbeck Network-Aware Automated Planning 24/75

Introduction Planning Agents Formalization Execution Agents Technical Approach Monitoring Agents Experiments Mixed-initiative UI Planning Agents Plan Evaluators: Steps. Agent Types: Alternatives. Domain-Independent. Longest temporally Random. ordered path. Guided. Duplicate plans. Department of Science Computer Kyle Usbeck Network-Aware Automated Planning 25/75

Introduction Planning Agents Formalization Execution Agents Technical Approach Monitoring Agents Experiments Mixed-initiative UI Planning Agents Plan Evaluators: Agent Types: (none). Domain-Independent. Random. Guided. Department of Science Computer Kyle Usbeck Network-Aware Automated Planning 25/75

Introduction Planning Agents Formalization Execution Agents Technical Approach Monitoring Agents Experiments Mixed-initiative UI Planning Agents Plan Evaluators: Agent Types: IED detection accuracy. Plan execution time. Domain-Independent. Network link quality. Random. Network bandwidth usage. Guided. Department of Science Computer Kyle Usbeck Network-Aware Automated Planning 25/75

Introduction Planning Agents Formalization Execution Agents Technical Approach Monitoring Agents Experiments Mixed-initiative UI Domain-Independent Planning Agent Uses I-Plan’s default strategy. I-Plan University of Edinburgh, Tate et al . ’s plan-space HTN planner which is built on an intelligent agent framework, I-X. Process Traverses search space depth-first. 1 Encounter an alternative whose constraints cannot be 2 satisfied. Backtracks using an A* search. 3 Department of Science Computer Kyle Usbeck Network-Aware Automated Planning 26/75

Introduction Planning Agents Formalization Execution Agents Technical Approach Monitoring Agents Experiments Mixed-initiative UI Domain-Independent Planning Agent Uses I-Plan’s default strategy. I-Plan University of Edinburgh, Tate et al . ’s plan-space HTN planner which is built on an intelligent agent framework, I-X. Process Traverses search space depth-first. 1 Encounter an alternative whose constraints cannot be 2 satisfied. Backtracks using an A* search. 3 Department of Science Computer Kyle Usbeck Network-Aware Automated Planning 26/75

Introduction Planning Agents Formalization Execution Agents Technical Approach Monitoring Agents Experiments Mixed-initiative UI Domain-Independent Planning Agent Uses I-Plan’s default strategy. I-Plan University of Edinburgh, Tate et al . ’s plan-space HTN planner which is built on an intelligent agent framework, I-X. Process Traverses search space depth-first. 1 Encounter an alternative whose constraints cannot be 2 satisfied. Backtracks using an A* search. 3 Department of Science Computer Kyle Usbeck Network-Aware Automated Planning 26/75

Introduction Planning Agents Formalization Execution Agents Technical Approach Monitoring Agents Experiments Mixed-initiative UI Domain-Independent Planning Agent Uses I-Plan’s default strategy. I-Plan University of Edinburgh, Tate et al . ’s plan-space HTN planner which is built on an intelligent agent framework, I-X. Process Traverses search space depth-first. 1 Encounter an alternative whose constraints cannot be 2 satisfied. Backtracks using an A* search. 3 Department of Science Computer Kyle Usbeck Network-Aware Automated Planning 26/75

Introduction Planning Agents Formalization Execution Agents Technical Approach Monitoring Agents Experiments Mixed-initiative UI Domain-Independent Planning Agent Uses I-Plan’s default strategy. I-Plan University of Edinburgh, Tate et al . ’s plan-space HTN planner which is built on an intelligent agent framework, I-X. Process Traverses search space depth-first. 1 Encounter an alternative whose constraints cannot be 2 satisfied. Backtracks using an A* search. 3 Department of Science Computer Kyle Usbeck Network-Aware Automated Planning 26/75

Introduction Planning Agents Formalization Execution Agents Technical Approach Monitoring Agents Experiments Mixed-initiative UI Random Planning Agent DFS with random branching. Process C ONSTRUCT R ANDOM P LAN ( I P ) 1: toVisit.push ( s 0 ) 2: while ¬ toVisit.empty () ∧ ¬ solution ( toVisit.peek ()) do 3: v ← toVisit.pop () 4: if v / ∈ visited then 5: visited.add ( v ) 6: r ← randomize ( v .children ()) 7: toVisit.push ( r ) 8: end if 9: end while 10: return toVisit.peek () Department of Science Computer Kyle Usbeck Network-Aware Automated Planning 27/75

Introduction Planning Agents Formalization Execution Agents Technical Approach Monitoring Agents Experiments Mixed-initiative UI Guided Planning Agent Generates qualitatively-different plans over: Domain-dependent criteria, and Network-centric criteria. Process A priority queue exists for each evaluator. 1 Every partial-plan is evaluated by all evaluators and placed 2 in their respective priority queues. The partial-plan at the head of each priority queue is used 3 for the next step. Department of Science Computer Kyle Usbeck Network-Aware Automated Planning 28/75

Introduction Planning Agents Formalization Execution Agents Technical Approach Monitoring Agents Experiments Mixed-initiative UI Guided Planning Agent Generates qualitatively-different plans over: Domain-dependent criteria, and Network-centric criteria. Process A priority queue exists for each evaluator. 1 Every partial-plan is evaluated by all evaluators and placed 2 in their respective priority queues. The partial-plan at the head of each priority queue is used 3 for the next step. Department of Science Computer Kyle Usbeck Network-Aware Automated Planning 28/75

Introduction Planning Agents Formalization Execution Agents Technical Approach Monitoring Agents Experiments Mixed-initiative UI Guided Planning Agent Generates qualitatively-different plans over: Domain-dependent criteria, and Network-centric criteria. Process A priority queue exists for each evaluator. 1 Every partial-plan is evaluated by all evaluators and placed 2 in their respective priority queues. The partial-plan at the head of each priority queue is used 3 for the next step. Department of Science Computer Kyle Usbeck Network-Aware Automated Planning 28/75

Introduction Planning Agents Formalization Execution Agents Technical Approach Monitoring Agents Experiments Mixed-initiative UI Guided Planning Agent Generates qualitatively-different plans over: Domain-dependent criteria, and Network-centric criteria. Process A priority queue exists for each evaluator. 1 Every partial-plan is evaluated by all evaluators and placed 2 in their respective priority queues. The partial-plan at the head of each priority queue is used 3 for the next step. Department of Science Computer Kyle Usbeck Network-Aware Automated Planning 28/75

Introduction Planning Agents Formalization Execution Agents Technical Approach Monitoring Agents Experiments Mixed-initiative UI Guided Planning Agent Planner new partial-plan Evaluator 1 Evaluator 2 Evaluator 3 Evaluator 4 Priority Queue Priority Queue Priority Queue Priority Queue Department of Science Computer Kyle Usbeck Network-Aware Automated Planning 29/75

Introduction Planning Agents Formalization Execution Agents Technical Approach Monitoring Agents Experiments Mixed-initiative UI Guided Planning Agent Planner next partial-plan evaluations Evaluator 1 Evaluator 2 Evaluator 3 Evaluator 4 Priority Queue Priority Queue Priority Queue Priority Queue Department of Science Computer Kyle Usbeck Network-Aware Automated Planning 30/75

Introduction Planning Agents Formalization Execution Agents Technical Approach Monitoring Agents Experiments Mixed-initiative UI Guided Planning Agent Planner new partial-plan Evaluator 1 Evaluator 2 Evaluator 3 Evaluator 4 Priority Queue Priority Queue Priority Queue Priority Queue Department of Science Computer Kyle Usbeck Network-Aware Automated Planning 31/75

Introduction Planning Agents Formalization Execution Agents Technical Approach Monitoring Agents Experiments Mixed-initiative UI Outline 1 Introduction Motivation Background Approach 2 Formalization Problem Statement 3 Technical Approach Planning Agents Execution Agents Monitoring Agents Mixed-initiative UI 4 Experiments Plan Evaluation Benchmarking Network-Aware Agent Combinations Discussion Department of Science Computer Kyle Usbeck Network-Aware Automated Planning 32/75

Introduction Planning Agents Formalization Execution Agents Technical Approach Monitoring Agents Experiments Mixed-initiative UI Execution Agents Service calls PI Plan Selection pI Execution Agent Faults Agent types: Defined by: Naïve. Service invocation. Reactive. Error handling. Proactive. Department of Science Computer Kyle Usbeck Network-Aware Automated Planning 33/75

Introduction Planning Agents Formalization Execution Agents Technical Approach Monitoring Agents Experiments Mixed-initiative UI Execution Agents Service calls PI Plan Selection pI Execution Agent Faults Agent types: Defined by: Naïve. Service invocation. Reactive. Error handling. Proactive. Department of Science Computer Kyle Usbeck Network-Aware Automated Planning 33/75

Introduction Planning Agents Formalization Execution Agents Technical Approach Monitoring Agents Experiments Mixed-initiative UI Naïve Execution Agent Naïve Execution Agent Properties Service Invocation Invokes services exactly as described by p I . The naïve agent requires that ∀ actions a ∈ p I , host ( a ) � = ∅ ∧ resources ( a ) � = {} . Error Handling Ignores execution errors. Not network-aware. Department of Science Computer Kyle Usbeck Network-Aware Automated Planning 34/75

Introduction Planning Agents Formalization Execution Agents Technical Approach Monitoring Agents Experiments Mixed-initiative UI Naïve Execution Agent Naïve Execution Agent Properties Service Invocation Invokes services exactly as described by p I . The naïve agent requires that ∀ actions a ∈ p I , host ( a ) � = ∅ ∧ resources ( a ) � = {} . Error Handling Ignores execution errors. Not network-aware. Department of Science Computer Kyle Usbeck Network-Aware Automated Planning 34/75

Introduction Planning Agents Formalization Execution Agents Technical Approach Monitoring Agents Experiments Mixed-initiative UI Naïve Execution Agent Naïve Execution Agent Properties Service Invocation Invokes services exactly as described by p I . The naïve agent requires that ∀ actions a ∈ p I , host ( a ) � = ∅ ∧ resources ( a ) � = {} . Error Handling Ignores execution errors. Not network-aware. Department of Science Computer Kyle Usbeck Network-Aware Automated Planning 34/75

Introduction Planning Agents Formalization Execution Agents Technical Approach Monitoring Agents Experiments Mixed-initiative UI Reactive Execution Agent Reactive Execution Agent Properties Service Invocation Invokes services exactly as described by p I . The reactive agent requires that ∀ actions a ∈ p I , host ( a ) � = ∅ ∧ resources ( a ) � = {} . Error Handling Repairs the failed p I by replacing failed service call(s) with new ones, creating p ′ I . Network-aware recovery — plan repair. Uses routing protocol neighbors & link quality. Department of Science Computer Kyle Usbeck Network-Aware Automated Planning 35/75

Introduction Planning Agents Formalization Execution Agents Technical Approach Monitoring Agents Experiments Mixed-initiative UI Reactive Execution Agent Reactive Execution Agent Properties Service Invocation Invokes services exactly as described by p I . The reactive agent requires that ∀ actions a ∈ p I , host ( a ) � = ∅ ∧ resources ( a ) � = {} . Error Handling Repairs the failed p I by replacing failed service call(s) with new ones, creating p ′ I . Network-aware recovery — plan repair. Uses routing protocol neighbors & link quality. Department of Science Computer Kyle Usbeck Network-Aware Automated Planning 35/75

Introduction Planning Agents Formalization Execution Agents Technical Approach Monitoring Agents Experiments Mixed-initiative UI Reactive Execution Agent Reactive Execution Agent Properties Service Invocation Invokes services exactly as described by p I . The reactive agent requires that ∀ actions a ∈ p I , host ( a ) � = ∅ ∧ resources ( a ) � = {} . Error Handling Repairs the failed p I by replacing failed service call(s) with new ones, creating p ′ I . Network-aware recovery — plan repair. Uses routing protocol neighbors & link quality. Department of Science Computer Kyle Usbeck Network-Aware Automated Planning 35/75

Introduction Planning Agents Formalization Execution Agents Technical Approach Monitoring Agents Experiments Mixed-initiative UI Reactive Execution Agent Reactive Execution Agent Properties Service Invocation Invokes services exactly as described by p I . The reactive agent requires that ∀ actions a ∈ p I , host ( a ) � = ∅ ∧ resources ( a ) � = {} . Error Handling Repairs the failed p I by replacing failed service call(s) with new ones, creating p ′ I . Network-aware recovery — plan repair. Uses routing protocol neighbors & link quality. Department of Science Computer Kyle Usbeck Network-Aware Automated Planning 35/75

Introduction Planning Agents Formalization Execution Agents Technical Approach Monitoring Agents Experiments Mixed-initiative UI Reactive Execution Agent Start Accept Plan no has more yes Execute Next Success actions? Action no failed? yes yes repaired? Repair Plan no Department of Science Computer Failure Kyle Usbeck Network-Aware Automated Planning 36/75

Introduction Planning Agents Formalization Execution Agents Technical Approach Monitoring Agents Experiments Mixed-initiative UI Proactive Execution Agent Proactive Execution Agent Properties Service Invocation Invokes services using network-aware logic to choose the host and resources at execution time. The proactive execution agent uses only service descriptions from actions a ∈ p I , meaning ∀ a ∈ p I , host ( a ) = ∅ ∧ resources ( a ) = {} Error Handling Repairs the failed p I by replacing failed service call(s) with new ones, creating p ′ I . Network-aware host/resource grounding. Department of Science Computer Kyle Usbeck Network-Aware Automated Planning 37/75

Introduction Planning Agents Formalization Execution Agents Technical Approach Monitoring Agents Experiments Mixed-initiative UI Proactive Execution Agent Proactive Execution Agent Properties Service Invocation Invokes services using network-aware logic to choose the host and resources at execution time. The proactive execution agent uses only service descriptions from actions a ∈ p I , meaning ∀ a ∈ p I , host ( a ) = ∅ ∧ resources ( a ) = {} Error Handling Repairs the failed p I by replacing failed service call(s) with new ones, creating p ′ I . Network-aware host/resource grounding. Department of Science Computer Kyle Usbeck Network-Aware Automated Planning 37/75

Introduction Planning Agents Formalization Execution Agents Technical Approach Monitoring Agents Experiments Mixed-initiative UI Proactive Execution Agent Proactive Execution Agent Properties Service Invocation Invokes services using network-aware logic to choose the host and resources at execution time. The proactive execution agent uses only service descriptions from actions a ∈ p I , meaning ∀ a ∈ p I , host ( a ) = ∅ ∧ resources ( a ) = {} Error Handling Repairs the failed p I by replacing failed service call(s) with new ones, creating p ′ I . Network-aware host/resource grounding. Department of Science Computer Kyle Usbeck Network-Aware Automated Planning 37/75

Introduction Planning Agents Formalization Execution Agents Technical Approach Monitoring Agents Experiments Mixed-initiative UI Proactive Execution Agent Start Accept Unground Plan no yes has more Ground First Plan Success actions? Action Execute Ground Plan Action no failed? yes yes repaired? Repair Plan no Department of Science Computer Failure Kyle Usbeck Network-Aware Automated Planning 38/75

Introduction Planning Agents Formalization Execution Agents Technical Approach Monitoring Agents Experiments Mixed-initiative UI Outline 1 Introduction Motivation Background Approach 2 Formalization Problem Statement 3 Technical Approach Planning Agents Execution Agents Monitoring Agents Mixed-initiative UI 4 Experiments Plan Evaluation Benchmarking Network-Aware Agent Combinations Discussion Department of Science Computer Kyle Usbeck Network-Aware Automated Planning 39/75

Introduction Planning Agents Formalization Execution Agents Technical Approach Monitoring Agents Experiments Mixed-initiative UI Monitoring Agents Methods of FDI Analytic. 1 Data-driven. 2 Knowledge-based. 3 Department of Science Computer Kyle Usbeck Network-Aware Automated Planning 40/75

Introduction Planning Agents Formalization Execution Agents Technical Approach Monitoring Agents Experiments Mixed-initiative UI Monitoring Agents Methods of FDI Analytic. ← Active Monitor 1 Data-driven. ← Passive Monitor 2 Knowledge-based. 3 Department of Science Computer Kyle Usbeck Network-Aware Automated Planning 40/75

Introduction Planning Agents Formalization Execution Agents Technical Approach Monitoring Agents Experiments Mixed-initiative UI Analytic Monitoring Agent Given the ordered plan p I = { a 0 , a 1 , . . . , a | p I | } An analytic monitoring agent: Constructs p M = { m 0 , m 1 , . . . , m | p I | + 1 } , an ordered set of 1 monitoring actions; I = � n Creates the new execution plan p ′ i = 0 { m i , a i } ; 2 The result is p ′ I = { m 0 , a 0 , m 1 , a 1 , . . . , m | p I | , a | p I | , m | p I | + 1 } . 3 Each m ∈ p M calculates the residual between expected 4 and actual bytes transferred. Department of Science Computer Kyle Usbeck Network-Aware Automated Planning 41/75

Introduction Planning Agents Formalization Execution Agents Technical Approach Monitoring Agents Experiments Mixed-initiative UI Analytic Monitoring Agent Given the ordered plan p I = { a 0 , a 1 , . . . , a | p I | } An analytic monitoring agent: Constructs p M = { m 0 , m 1 , . . . , m | p I | + 1 } , an ordered set of 1 monitoring actions; I = � n Creates the new execution plan p ′ i = 0 { m i , a i } ; 2 The result is p ′ I = { m 0 , a 0 , m 1 , a 1 , . . . , m | p I | , a | p I | , m | p I | + 1 } . 3 Each m ∈ p M calculates the residual between expected 4 and actual bytes transferred. Department of Science Computer Kyle Usbeck Network-Aware Automated Planning 41/75

Introduction Planning Agents Formalization Execution Agents Technical Approach Monitoring Agents Experiments Mixed-initiative UI Analytic Monitoring Agent Given the ordered plan p I = { a 0 , a 1 , . . . , a | p I | } An analytic monitoring agent: Constructs p M = { m 0 , m 1 , . . . , m | p I | + 1 } , an ordered set of 1 monitoring actions; I = � n Creates the new execution plan p ′ i = 0 { m i , a i } ; 2 The result is p ′ I = { m 0 , a 0 , m 1 , a 1 , . . . , m | p I | , a | p I | , m | p I | + 1 } . 3 Each m ∈ p M calculates the residual between expected 4 and actual bytes transferred. Department of Science Computer Kyle Usbeck Network-Aware Automated Planning 41/75