01

Understand the structure of the problem

Two instances with the same job count can need different search methods. Time-window slack, skill scarcity, resource substitutability, geography, prerequisites, and work already in progress all affect feasibility and the value of search.

A lost resource, a tight time window, and a prerequisite motivate a testable search hypothesis. This teaching example has no solver run.Business state and constraintsTeaching example / Resource change at 09:00Unit AElectrical EUnit BUnavailableUnit CGeneral workJ1 active · Free at 09:20J6 equipment loadedJ2 / Cert. E09:40–09:50J3 / Cert. E10:15–10:45J4 / PrerequisiteThenJ1 stays lockedQualifications and execution facts constrain adjustmentsEvidenceCompeting for skillsOnly A qualifies for J2 and J3Time slack09:20 + 24min → 09:44Dependency across crewsJ3 waits for J4Locked commitmentsWork already in progress, fixed assignments, and actual loadsJ2 starts by 09:50 · 6 min slackCandidate search programsGenerate · Select · Allocate budgetA Existing strategy and operatorsImplementedWarm start → Nearest candidates → Local searchrelocate / swap / Or-opt · Late AcceptanceB Design search for the bottleneckResearch candidateScarcity-firstLocal refinementRemove / RepairRefine againHypothesis: reserve scarce resources, then repair linked jobs togetherRelated removal, regret-k, and control need implementation and testsSolving and independent full scoring use all execution facts and the original directed matrix.Constraints, observed bottlenecks, and candidate programs. The teaching hypothesis has no solver run.Business state and constraintsTeaching example / Resource change at 09:00Unit AElectrical EUnit BUnavailableUnit CGeneral workJ1 active · Free at 09:20J6 equipment loadedJ2 / Cert. E09:40–09:50J3 / Cert. E10:15–10:45J4 / PrerequisiteThenJ1 stays lockedQualifications and execution facts constrain adjustmentsFind bottlenecks in constraintsCompeting for skillsOnly A qualifies for J2 and J3Time slack09:20 + 24min → 09:44Dependency across crewsJ3 waits for J4Locked commitmentsWork already in progress, fixed assignments, and actual loadsJ2 latest start 09:506 min slack on direct routeCandidate search programsA Existing strategy and operatorsImplementedWarm start → Nearest candidates→ Local searchrelocate / swap / Or-opt · Late AcceptanceB Design search for the bottleneckResearch candidates · Generate, select, budget01 Scarcity-first02 Refine03 Remove / Repair04 Refine againHypothesis: reserve scarce resources, then repair linked jobsTest related removal, regret-k, and controlFull facts feed solving and independent scoringKeep the original directed matrix alongside the summary.

Start with explicit features and relationships. Add learned representations or change state encodings when downstream gains justify the computing cost.

02

A method specifies how to build an initial solution, prioritize moves, combine neighborhoods, perturb and repair plans, and allocate the budget across stages. LLMs can propose hypotheses and candidate programs for isolated runs and scenario benchmarks.

ParticipantRoleCurrent scope
Optimization engineRun search, score results, and explain constraintsImplemented in vrp-1. Performance varies by instance, budget, and environment.
LLMs and agentsInterpret requirements, propose candidates, coordinate tools, and explain resultsMethod generation and workflow search still need dedicated validation.
Rules, specialized models, and short trialsPredict candidate performance, shortlist, and allocate budgetLearned predictions and Jev-style semantic assessment are research topics.
Operations teamsConfirm key business rules, trade-offs, and adoptionBusiness objectives require human approval.

03

vrp-1 solver benchmark

The first public synthetic benchmark compares vrp-0 and vrp-1 on six shared scenarios. It runs cold construction and continued optimization from identical starting routes separately. Each run uses one solver thread and a 30-second budget, with five random seeds per group.

Cold starts · Six synthetic benchmarksCold startsvrp-1vrp-0Driving hours ↓ / Solve time (s) with setup · Separate y-scalesa01 · Basic delivery50 jobs / 10 vehicles−14.95%1.51.7522.2501530a02 · Dense urban200 jobs / 8 vehicles−31.82%11.522.501530a03 · Four depots480 jobs / 80 vehicles−7.14%1517.52001530a04 · Larger scale1,000 jobs / 100 vehicles−31.93%10012515001530a05 · Skills and capacity200 jobs / 20 vehicles−2.82%67801530a06 · Tight windows50 jobs / 4 vehicles−6.54%22.501530Cold starts · Six synthetic benchmarksCold startsvrp-1vrp-0Driving hours ↓ / Solve time (s) · Separate y-scalesa01 · Basic delivery50 jobs / 10 vehicles−14.95%1.5201530a02 · Dense urban200 jobs / 8 vehicles−31.82%1201530a03 · Four depots480 jobs / 80 vehicles−7.14%1517.52001530a04 · Larger scale1,000 jobs / 100 vehicles−31.93%10015001530a05 · Skills and capacity200 jobs / 20 vehicles−2.82%6801530a06 · Tight windows50 jobs / 4 vehicles−6.54%22.501530
Continue from identical routes · Six synthetic benchmarksContinue from identical routesvrp-1vrp-0Driving hours ↓ / Solve time (s) with setup · Separate y-scalesa01 · Basic delivery50 jobs / 10 vehicles−8.08%23401530a02 · Dense urban200 jobs / 8 vehicles−26.51%1234501530a03 · Four depots480 jobs / 80 vehicles−9.72%20304001530a04 · Larger scale1,000 jobs / 100 vehicles−17.96%10012515017501530a05 · Skills and capacity200 jobs / 20 vehicles−5.03%57.51012.51501530a06 · Tight windows50 jobs / 4 vehicles−11.50%22.5301530Continue from identical routes · Six synthetic benchmarksContinue from identical routesvrp-1vrp-0Driving hours ↓ / Solve time (s) · Separate y-scalesa01 · Basic delivery50 jobs / 10 vehicles−8.08%23401530a02 · Dense urban200 jobs / 8 vehicles−26.51%2401530a03 · Four depots480 jobs / 80 vehicles−9.72%20304001530a04 · Larger scale1,000 jobs / 100 vehicles−17.96%10015001530a05 · Skills and capacity200 jobs / 20 vehicles−5.03%5101501530a06 · Tight windows50 jobs / 4 vehicles−11.50%2301530

Curves use observed best-solution events. Total driving hours appear once all five seeds have fully assigned plans with zero hard violations. Shading shows the min–max range. Time includes preprocessing and construction. Each plot has its own vertical scale. Percentages are the median paired changes in driving time against vrp-0.

Charts are drawn from best-solution events and paired statistics in the first vrp-1 benchmark. The results apply to those versions, configurations, scenarios, and budgets. New methods are tested on representative scenarios and regression suites before human approval for promotion.

04

Online decisions and offline engine improvement

Two decision loops, intake and solver routing, control plane and engines, published assets and feedback, and external map servicesOnline decisionsAgents interpret, coordinate, and explainIntake routingWorkflow templates and published versionsInterpret scenario and adapt inputsJobs / Resources / Constraints / GoalsSolver routingStrategy / Engine version / BudgetApproval and validationBusiness rules / Data / AccessDecision RuntimeControl plane · Scheduling and resultsScenario solverExecution · Isolated optimizationExplain, compare, refineUnassigned jobs / Adjust / RerunApproval and execution feedbackAdopt / Replay / ExecuteOffline solver improvementAutomated generation and evolution are research topicsSamples and regression setsGoals / Exceptions / FeedbackCandidate models, algorithms, workflowsPropose / Generate / ReviseIsolated build and scenario testsSimulation / Benchmarks / Trade-offsApprove, promote, registerEngines / Algorithms / RoutingExternal map serviceRoutes / Distance matrixOutside the platformShared governanceAudit · Versions · Security · Rights · Quotas · Cost · LineageInfrastructure: Kubernetes / Network / Compute / Storage / Images / TelemetryOnline decision and offline improvement loops with feedback, published assets, external maps, and shared governanceOnline decisionsIntake routingWorkflow templates and published versionsInterpret scenario and adapt inputsJobs / Resources / Constraints / GoalsSolver routingStrategy / Engine version / BudgetApproval and validationBusiness rules / Data / AccessDecision RuntimeControl plane · Scheduling and resultsScenario solverExecution · Isolated optimizationExplain, compare, refineUnassigned jobs / Adjust / RerunApproval and execution feedbackAdopt / Replay / ExecuteExternal map serviceRoutes / Distance matrixOutside the platformResults, diagnostics, and feedback → Samples and regressionOffline solver improvementAutomated generation and evolution are research topicsSamples and regression setsGoals / Exceptions / FeedbackCandidate models, algorithms, workflowsPropose / Generate / ReviseIsolated build and scenario testsSimulation / Benchmarks / Trade-offsApprove, promote, registerEngines / Algorithms / RoutingApproved assets feed intake and solver routingShared governanceAudit · Versions · Security · Data rightsQuotas · Cost · LineageShared infrastructureKubernetes / Network / Compute / StorageRegistry / Telemetry

Intake routing selects a published workflow. Solver routing selects the strategy, engine version, configuration, and budget. The runtime validates assets, pins immutable versions, and manages scheduling and results. Engines solve within resource limits and send results back for explanation, comparison, adjustment, and approval.

Online results and human feedback feed the scenario and regression sets. Validated offline assets become available to online routing after approval. Versioning, audit, security, authorization, cost, and quotas apply to both loops. Map and travel-matrix services are external dependencies. Compute and storage support the runtime.

05

Capability status and research

StatusDescription
Implementedvrp-1 core, workbench, isolated execution, business constraints, heuristic search, incremental scoring, independent full scoring, explanations, and execution replanning.
ValidatedEvidence with documented data, baselines, budgets, and scope. Historical cases and demos have their own labels.
In developmentPlatform-level experiments across alternatives and guided infeasibility adjustments are in development. Delivery scope depends on the version.
Roadmap and researchLearned problem representations, instance-level performance prediction, dynamic operator combinations, program generation, agent workflow search, and transfer across scenarios.

Existing components include greedy insertion, warm starts, relocate, swap, Or-opt, perturbation and repair, Late Acceptance, multiple starts, and candidate neighborhoods. Default configurations depend on benchmark results across instances and budgets.

Models propose candidates for deterministic computation, rules, simulations, real cases, and human acceptance to verify. Independent full scoring checks the returned plan. A global optimum requires a separate proof.

06

Working with existing approaches

ApproachStrengthsImplementation considerations
Manual planningFamiliar with operations and local exceptionsTrade-offs across operations, scaling, and repeatability
In-house developmentFit to internal processes and control over dataOptimization expertise, engineering, and long-term maintenance
Off-the-shelf platforms and GIS APIsMaps, routes, and standard capabilitiesMultiple resources, specialist constraints, and domain workflows
Specialized optimization softwareMature modeling and optimizationScenario adaptation, integration, and ongoing delivery
Custom implementation servicesTailored to a specific problemVersion reuse, ongoing operations, and expansion

PsiAI connects business modeling, optimization, and continuous verification. It can work alongside existing tools and infrastructure.

Start with a concrete planning problem.

Contact the PsiAI team ↗