Regulation

from hackb_judge import Problem, Evaluator  # optional, type-only

def solve(problem: Problem, evaluate: Evaluator) -> object:
    ...

Run pip install hackb-judge to get IDE autocomplete. The package is type-stub only — no runtime dependency.

• problemREQUIRED

  TYPE

  Problem (TypedDict)

  A TypedDict passed as the first argument to solve. It carries the keys below.
  EXAMPLE

  {
      "problem_id": "sphere-5d",
      "search_space": {
          "variables": [
              {"name": "x1", "type": "continuous", "lower": -5.12, "upper": 5.12},
              {"name": "x2", "type": "continuous", "lower": -5.12, "upper": 5.12},
          ],
      },
      "budget": 100,
      "seed": 42,
      "initial_data": [
          {"x": {"x1": 0.0, "x2": 0.0}, "y": 0.0},
      ],
  }

  • problem_idREQUIRED

    TYPE

    str

    Identifier of the problem. Available to your code for identification only.
    EXAMPLE

    "sphere-5d"

  • search_spaceREQUIRED

    TYPE

    SearchSpace

    Search-space definition. Each dimension has a name (key) and an admissible range (continuous: [low, high]; discrete: an enumeration). Every x dict you pass to evaluate must use this set of keys.
    EXAMPLE

    {
        "variables": [
            # 連続値: float, lower <= x <= upper
            {"name": "lr",        "type": "continuous",  "lower": 1e-5, "upper": 1.0},
            # 整数: int (両端含む)
            {"name": "layers",    "type": "integer",     "lower": 1,    "upper": 8},
            # カテゴリカル: choices からひとつ
            {"name": "optimizer", "type": "categorical", "choices": ["sgd", "adam"]},
            # バイナリ: 0 or 1
            {"name": "dropout",   "type": "binary"},
        ],
    }

  • budgetREQUIRED

    TYPE

    int

    Number of objective-function evaluations you may consume. Caps evaluate calls inside solve. Calls past the budget are rejected.
    EXAMPLE

    100

  • seedREQUIRED

    TYPE

    int

    Random seed used by the judge for scoring. If your code uses randomness, reusing this seed lets you reproduce a run.
    EXAMPLE

    1

  • initial_dataOPTIONAL

    TYPE

    list[InitialPoint]

    DEFAULT

    []

    Pre-labelled observations. Each entry is a dict with two keys: x (a dict that respects search_space) and y (the observed float). Some problems ship an empty list.
    EXAMPLE

    [
        {"x": {"x1": 0.5,  "x2": -0.3}, "y": 0.34},
        {"x": {"x1": -1.2, "x2":  0.8}, "y": 2.08},
    ]

• evaluateREQUIRED

  TYPE

  Callable[[dict[str, Any]], float]

  A callable that evaluates the unknown objective f one point at a time. Decorator-based libraries (e.g. amplify-bbopt) can wrap this callable as their black-box function directly.

  • x must contain every dimension declared in search_space. Each value must respect that dimension's type and range.
  • The return value is a finite scalar (no NaN, no ±inf). Observations may include observation noise.
  • Each call counts as one evaluation. Calls past budget raise ValueError.
  • If x violates the search space, the call raises ValueError. Your code may catch it and retry, but the call still counts toward budget.
  • The objective function f and ground-truth labels are not visible from the submission. They run outside the sandbox boundary inside the judge.
  EXAMPLE

  # 1 点ずつ評価。ジャッジ側で budget をカウントする。
  y = evaluate({"x1": 0.5, "x2": -0.3})
  # y は float (例: 0.34)。NaN / ±inf は返らない。
  
  # search_space 違反は ValueError として上がる (呼び出しは消費される)。
  try:
      evaluate({"x1": 99.0, "x2": 0.0})
  except ValueError:
      ...
  
  # budget 超過時も ValueError。
  for _ in range(budget + 1):
      evaluate(sample_x())  # 最後の 1 回が ValueError

The concrete search_space, budget, and initial_data values for each problem are visible from the problem's detail page once the contest opens. Only the invariant types and meanings are documented here.