{
  "field_id": "F036",
  "field": "combinatorial optimization",
  "directions": [
    {
      "direction_id": "G0",
      "name": "Direct generation of discrete solutions",
      "scope": "Directly form routes, graph solutions, orderings, or other discrete solutions through sequential choices, joint prediction or sampling, or simulated-lookahead tree search; these approaches need not occur together."
    },
    {
      "direction_id": "G1",
      "name": "Transfer of learned optimization heuristics",
      "scope": "Enable learned combinatorial optimization heuristics to remain useful beyond their original training setting, across different instance sizes, distributions, or related problems."
    },
    {
      "direction_id": "G2",
      "name": "Operational schedule construction and optimization",
      "scope": "Produce usable schedules under temporal, equipment, resource, and operational constraints, including through direct decisions by learned dispatch policies or by formulating constraint models for solving or targeted reoptimization."
    },
    {
      "direction_id": "G3",
      "name": "Iterative heuristic improvement",
      "scope": "Repeatedly search around an existing solution or a continually evolving population of candidates, improving solution quality through revisions, neighborhoods and repair, or control of a host metaheuristic’s behavior."
    },
    {
      "direction_id": "G4",
      "name": "Control of exact optimization procedures",
      "scope": "Make exact optimization more efficient by improving search-tree expansion, valid-inequality management, or iterative column generation."
    },
    {
      "direction_id": "G5",
      "name": "Decision-space reduction for downstream optimization",
      "scope": "Before handing a problem to a downstream solver or search process, predict assignments, eliminate input candidates, or select decisions to fix or release to focus subsequent optimization; this includes both initial solving and reoptimization based on an existing solution."
    },
    {
      "direction_id": "G6",
      "name": "Coordinated multi-agent pathfinding",
      "scope": "Find collision-free paths for multiple agents and study how task-goal assignment, search structure, encodings, or graph conditions affect solvability."
    },
    {
      "direction_id": "G7",
      "name": "Decision-focused prediction for optimization",
      "scope": "Train models according to the quality of decisions produced when predicted parameters enter downstream optimization, or construct learnable surrogates for that decision objective."
    },
    {
      "direction_id": "G8",
      "name": "Submodular selection algorithms",
      "scope": "Select sets under a submodular value objective while studying group-fairness requirements or algorithms and guarantees under limits on queries, data access, or computation; the two types of requirements need not occur together."
    },
    {
      "direction_id": "G9",
      "name": "Clustering and grouping optimization",
      "scope": "Optimize the clustering or grouping of entities so that the resulting groups meet demographic fairness and service-quality requirements, or align as closely as possible with agreement and disagreement relations among entities; neither objective is a prerequisite for the other."
    },
    {
      "direction_id": "G10",
      "name": "Facility siting for spatial service",
      "scope": "Choose facility locations or capacities to serve spatially distributed demand while accounting for accessibility, network costs, or service requirements."
    }
  ]
}