Interactive Resource Management in the COMIREM Planner Stephen

Interactive Resource Management in the COMIREM Planner Stephen F. Smith, David Hildum, David Crimm Intelligent Coordination and Logistics Lab The Robotics Institute Carnegie Mellon University Pittsburgh PA 15213 sfs@cs. cmu. edu 412 -268 -8811 Carnegie Mellon Workshop on Mixed-Initiative Intelligent Systems - August 9 2003 IJCAI-03

Outline of Talk – Brief Introduction to Comirem – Mixed-Initiative Perspective – Connection to Workshop Themes Carnegie Mellon

COMIREM A light-weight, interactive system for resource management in continuous planning domains Domain: SOF planning Motivating Themes: – Resource management cannot be considered a separable post-process to plan generation – Planning is an ill-structured, iterative process that is rarely amenable to total automation and not well supported by batch-oriented solution generators – Planning involves collaboration among (increasingly) mobile decision making agents Carnegie Mellon

Embassy Rescue Scenario Rebel Enclave Staging Area Task Force Charlie (56 Troops) Bridge Task Force Bravo (64 Troops) Task Force Alpha (24 Troops) Home Airport Carnegie Mellon Available at Home Airport – 7 MH 60 s – – 5 MH 47 s 5 MC-130 Hs 2 C-141 s 1 AC-130 U Rebel Controlled Airport Embassy 250 Am. Cits

Mixed-Initiative Design Goals Adjustable decision-making autonomy – User will want to make decisions at different levels of detail in different contexts Translation of system models and decisions – User should be able to inject decisions without having to understand system search models and vice-versa be able to effectively interpret system results Incremental problem solving – Constraints typically become known incrementally – Controlled change facilitates comprehension – Solution stability is crucial in continuous planning domains Carnegie Mellon

Constraint Management and Search Infra-structure Comirem is a flexible times scheduler: – Activity start and end times float to the extent that problem constraints allow – Activities requiring the same resources are sequenced Simple Temporal Problem (STP) constraint network solver is used to manage temporal constraints – Constraint graph of time points (nodes) and distances (arcs) Higher-level domain model super-imposed to add reasoning about resource usage constraints – Required and provided capabilities – Resource location (positioning, de-positioning, repositioning) – Resource carrying capacity (manifests and configurations) Decisions (user and system) are made opportunistically Carnegie Mellon

Elements of Mixed-Initiative Approach Highly interactive - spreadsheet metaphor Levels of automated decision-making – Individual decision expansion and options – Temporal and resource feasibility checking – Automated solution generation -biased by user goals and preferences – User over-ride of any constraint in system model Interaction via mutually understood domain model – Translation of domain model edits into internal constraint models – Complementary use of domain model to convey and interpret results Visualization of decision impact Carnegie Mellon

Carnegie Mellon

Generating Options Light-Transport-Activity MH-60 Capacity: 14 Resource Reqs. instance Deploy(A, B, ? Res) MH-60 -5 MH-60 -4 MH-60 -3 MH-60 -2 MH-60 -1 instance MH-47 Capacity: 40 Manifest: 120 MH-60 Option Carnegie Mellon MH-47 Option MH-60 -4 MH-60 -3 MH-60 -2 MH-47 -1

Generating Conflict Resolution Options LFT A Move 1 x. MH-60 B EST Dur(MH-60) > LFT(Move) - EST(Move) MH-60 Nom. Speed: 150 Option 4: Deploy earlier Airdrop-Activity TF-Deploy-Time CZ Carnegie Mellon C 130 Nom. Speed: 200 Start. Time-Constraint instance Res reqs. Option 1: Override computed duration Option 2: Assign a faster resource instance Move(A, B, MH-60) Due. Date-Constraint instance <0, dd. Move> Detected Cycle Magnitude: m TF-Engage-Time Option 3: Delay engagement

Comirem Positions on Workshop Issues Task - User manipulates goals, constraints and preferences; system solves within specified parameters Control - User in control; system offers decision options wherever possible and solutions when user delegates Awareness - Mutually understandable domain model used to bridge user and system models Communication - Summarization, visualization of decision impact Evaluation - increased efficiency/effectiveness; system manages complexity; user brings knowledge outside of system models Architecture - Spreadsheet model of interaction; incremental change Carnegie Mellon

END Carnegie Mellon

Functional Capabilities Interactive Planning and Resource Allocation – – – Option generation Visualization of decision impact Requirements and capabilities editing Automated assignment and feasibility checking What-if exploration Resource Configuration – Construction and allocation of aggregate resources Execution Management – Resource tracking – Plan tracking – Conflict analysis Carnegie Mellon

A More Complex Conflict Involving Shared Resources C Move 2 1 x. HMMVV B A Move 1 1 x. HMMVV LFT EST Dur(HMMVV) > LFT(Move) - EST(Move) • Resource sequencing constraint in conjunction with the timing constraints of Move 1 and Move 2 causes “cycle” Carnegie Mellon

Gantt and Vector Activity Views Comirem User Interface Resource Aggregation Resource Usage & Positioning Carnegie Mellon Resource Tracking

Carnegie Mellon