AGENT SKILL

map-optimization-strategy

Strategy for solving constraint optimization problems on spatial maps. Use when you need to place items on a grid/map to maximize some objective while satisfying constraints.

License 许可
Apache-2.0
最近更新
2026-11-14
Stars
★ 0
Forks
1

01 前言 Front Matter

name
map-optimization-strategy
description
Strategy for solving constraint optimization problems on spatial maps. Use when you need to place items on a grid/map to maximize some objective while satisfying constraints.

03 SKILL.md 正文

Map-Based Constraint Optimization Strategy

A systematic approach to solving placement optimization problems on spatial maps. This applies to any problem where you must place items on a grid to maximize an objective while respecting placement constraints.

Why Exhaustive Search Fails

Exhaustive search (brute-force enumeration of all possible placements) is the worst approach:

  • Combinatorial explosion: Placing N items on M valid tiles = O(M^N) combinations
  • Even small maps become intractable (e.g., 50 tiles, 5 items = 312 million combinations)
  • Most combinations are clearly suboptimal or invalid

The Three-Phase Strategy

Phase 1: Prune the Search Space

Goal: Eliminate tiles that cannot contribute to a good solution.

Remove tiles that are:

  1. Invalid for any placement - Violate hard constraints (wrong terrain, out of range, blocked)
  2. Dominated - Another tile is strictly better in all respects
  3. Isolated - Too far from other valid tiles to form useful clusters
Before: 100 tiles in consideration
After pruning: 20-30 candidate tiles

This alone can reduce search space by 70-90%.

Phase 2: Identify High-Value Spots

Goal: Find tiles that offer exceptional value for your objective.

Score each remaining tile by:

  1. Intrinsic value - What does this tile contribute on its own?
  2. Adjacency potential - What bonuses from neighboring tiles?
  3. Cluster potential - Can this tile anchor a high-value group?

Rank tiles and identify the top candidates. These are your priority tiles - any good solution likely includes several of them.

Example scoring:
- Tile A: +4 base, +3 adjacency potential = 7 points (HIGH)
- Tile B: +1 base, +1 adjacency potential = 2 points (LOW)

Phase 3: Anchor Point Search

Goal: Find placements that capture as many high-value spots as possible.

  1. Select anchor candidates - Tiles that enable access to multiple high-value spots
  2. Expand from anchors - Greedily add placements that maximize marginal value
  3. Validate constraints - Ensure all placements satisfy requirements
  4. Local search - Try swapping/moving placements to improve the solution

For problems with a "center" constraint (e.g., all placements within range of a central point):

  • The anchor IS the center - try different center positions
  • For each center, the reachable high-value tiles are fixed
  • Optimize placement within each center's reach

Make Phase 3 Operational (bounded search loop)

When maximizing an objective (not just finding any valid placement), do not stop after validating a single plausible configuration. Instead, run a bounded anchor search plus a small local-improvement loop.

  1. Enumerate anchor candidates (centers) in a bounded way
  • Build a set of feasible anchors from Phase 1 constraints.
  • If this set is large, restrict to top-K anchors using a cheap proxy score (e.g., count of high-value tiles within the allowed radius, or sum of their tile scores).
  1. For each anchor, solve placements with a greedy pass
  • Fix the anchor.
  • Recompute the candidate set within the anchor’s allowed range.
  • Place items one-by-one by marginal gain (objective delta) while enforcing hard constraints after each addition (constraint propagation).
  1. Local improvement (small neighborhood search) Perform a limited number of improvement attempts so runtime stays predictable:
  • Consider single-tile moves for one placed item to another candidate tile.
  • Consider pairwise swaps of two placed items.
  • Keep a change only if it improves the objective and remains valid.

Suggested bounds (tune to map size):

  • K_anchors = 10–30
  • max_local_iters = 50–200
  • max_neighbors_per_item = 10–30 (e.g., evaluate only the top-N alternative tiles by heuristic)
  1. Track best-so-far with a strict validator Maintain best_solution across anchors and local steps. Every time you accept a new best:
  • Re-validate all placements against the hard constraints.
  • Recompute the objective from the authoritative scoring rules.

Fallback/verification step: If the search budget is exhausted, return the best validated solution found so far (never an unvalidated intermediate). If results are unexpectedly low or unstable, increase K slightly or widen the local neighborhood, but keep the search bounded.

Algorithm Skeleton

def optimize_placements(map_tiles, constraints, num_placements):
    # Phase 1: Prune
    candidates = [t for t in map_tiles if is_valid_tile(t, constraints)]

    # Phase 2: Score and rank
    scored = [(tile, score_tile(tile, candidates)) for tile in candidates]
    scored.sort(key=lambda x: -x[1])  # Descending by score
    high_value = scored[:top_k]

    # Phase 3: Anchor search
    best_solution = None
    best_score = 0

    for anchor in get_anchor_candidates(high_value, constraints):
        solution = greedy_expand(anchor, candidates, num_placements, constraints)
        solution = local_search(solution, candidates, constraints)

        if solution.score > best_score:
            best_solution = solution
            best_score = solution.score

    return best_solution

Key Insights

  1. Prune early, prune aggressively - Every tile removed saves exponential work later

  2. High-value tiles cluster - Good placements tend to be near other good placements (adjacency bonuses compound)

  3. Anchors constrain the search - Once you fix an anchor, many other decisions follow logically

  4. Greedy + local search is often sufficient - You don't need the global optimum; a good local optimum found quickly beats a perfect solution found slowly

  5. Constraint propagation - When you place one item, update what's valid for remaining items immediately

Common Pitfalls

  • Ignoring interactions - Placing item A may change the value of placing item B (adjacency effects, mutual exclusion)

  • Over-optimizing one metric - Balance intrinsic value with flexibility for remaining placements

  • Forgetting to validate - Always verify final solution satisfies ALL constraints

  • Single-shot confirmation - Validating one hand-picked layout is not an optimization strategy. Use a bounded anchor loop + local improvements to reliably find a better (or best-found) solution under a fixed time budget.

  • Shell heredoc + chaining mistakes - If your executor/tool has a single-command constraint, either (a) write the output in one command and verify in a second command, or (b) chain only with operators that are part of the same shell command. When using a heredoc (<<'PY'), ensure the terminator line ends the command cleanly—do not append extra tokens on the terminator line.

04 同仓库更多 Skills