"""Sensitivity analysis of BudgetOptimizer heuristic weights.

Scientifically tests whether the hand-set BUDGET_PRIORITY_DEFAULTS weights are
"load-bearing" for the system's user-facing decisions, by:

  (A) Showing the tier-feasibility decision (check_budget_feasibility) is
      INVARIANT to weight perturbation (it reads empirical per-destination
      cost_tiers data, not the weights).
  (B) Monte-Carlo perturbing the weights (each default x U[0.7, 1.3]) and
      measuring how much optimize_with_priorities outputs move + whether the
      recommended accommodation/food tiers flip.

Run:  python -m scripts.sensitivity_analysis   (from chatbot-ml-service/)
"""

import json
import os
import random
import sys

sys.path.insert(0, os.path.dirname(os.path.dirname(os.path.abspath(__file__))))

from app.models.budget_optimizer.optimizer import BudgetOptimizer  # noqa: E402

KB_PATH = os.path.join(
    os.path.dirname(os.path.dirname(os.path.abspath(__file__))),
    "app", "data", "knowledge_base", "destinations.json",
)
DEST_IDS = ["da-nang", "phu-quoc", "hoi-an"]
SCENARIOS = [
    # (budget_vnd, days, adults, children, month)
    (5_000_000, 3, 2, 1, 7),
    (8_000_000, 4, 2, 0, 3),
    (15_000_000, 5, 2, 2, 12),
    (3_000_000, 3, 1, 0, None),
]
USER_PRIORITIES = {"accommodation": "high", "food": "low", "activities": "medium"}
N_DRAWS = 500
PERTURB_LOW, PERTURB_HIGH = 0.70, 1.30  # +/-30% per-category noise


def _perturbed_weights(base: dict, rng: random.Random) -> dict:
    return {k: v * rng.uniform(PERTURB_LOW, PERTURB_HIGH) for k, v in base.items()}


def _max_rel_delta(a: dict, b: dict) -> float:
    out = 0.0
    for k in a:
        denom = max(1.0, abs(a[k]))
        out = max(out, abs(a[k] - b[k]) / denom)
    return out


def main() -> None:
    with open(KB_PATH, encoding="utf-8-sig") as f:
        dests = json.load(f)["destinations"]
    opt = BudgetOptimizer(dests)
    base = dict(BudgetOptimizer.BUDGET_PRIORITY_DEFAULTS)
    rng = random.Random(42)

    print("=" * 78)
    print("SENSITIVITY ANALYSIS - BudgetOptimizer heuristic weights")
    print(f"base weights: {base}")
    print(f"Monte-Carlo: {N_DRAWS} draws, each weight x U[{PERTURB_LOW},{PERTURB_HIGH}]")
    print("=" * 78)

    # ---- (A) tier-feasibility invariance ------------------------------------
    print("\n[A] check_budget_feasibility() - core user-facing decision")
    print(f"{'destination':<10}{'budget':>11}{'days':>5}{'pax':>5}"
          f"{'tier(base)':>13}{'tier flips/' + str(N_DRAWS):>16}")
    total_flips_A = 0
    for did in DEST_IDS:
        dest = opt.get_destination(did)
        for budget, days, ad, ch, month in SCENARIOS:
            base_res = opt.check_budget_feasibility(dest, budget, "VND", days, ad, ch, month)
            base_tier = base_res.get("recommended_tier") or "infeasible"
            flips = 0
            for _ in range(N_DRAWS):
                opt.BUDGET_PRIORITY_DEFAULTS = _perturbed_weights(base, rng)
                r = opt.check_budget_feasibility(dest, budget, "VND", days, ad, ch, month)
                if (r.get("recommended_tier") or "infeasible") != base_tier:
                    flips += 1
            opt.BUDGET_PRIORITY_DEFAULTS = base
            total_flips_A += flips
            print(f"{did:<10}{budget:>11,}{days:>5}{ad + ch:>5}"
                  f"{base_tier:>13}{flips:>16}")
    print(f"  => tier-feasibility flips across all cases: {total_flips_A} "
          f"(expected 0 - decision uses empirical cost_tiers, NOT weights)")

    # ---- (B) optimize_with_priorities allocation sensitivity ----------------
    print("\n[B] optimize_with_priorities() - advisory allocation (DOES use weights)")
    print(f"user priorities: {USER_PRIORITIES}")
    print(f"{'destination':<10}{'budget':>11}{'days':>5}"
          f"{'accomTier flips':>16}{'foodTier flips':>16}{'maxRelDelta(mean)':>20}")
    worst_delta = 0.0
    total_flips_B = 0
    for did in DEST_IDS:
        dest = opt.get_destination(did)
        for budget, days, ad, ch, month in SCENARIOS:
            base_res = opt.optimize_with_priorities(
                dest, budget, "VND", days, ad, ch, USER_PRIORITIES, month)
            base_alloc = base_res["allocation"]
            base_tiers = base_res["recommended_tiers"]
            a_flip = fd_flip = 0
            deltas = []
            for _ in range(N_DRAWS):
                opt.BUDGET_PRIORITY_DEFAULTS = _perturbed_weights(base, rng)
                r = opt.optimize_with_priorities(
                    dest, budget, "VND", days, ad, ch, USER_PRIORITIES, month)
                if r["recommended_tiers"]["accommodation"] != base_tiers["accommodation"]:
                    a_flip += 1
                if r["recommended_tiers"]["food"] != base_tiers["food"]:
                    fd_flip += 1
                deltas.append(_max_rel_delta(base_alloc, r["allocation"]))
            opt.BUDGET_PRIORITY_DEFAULTS = base
            mean_delta = sum(deltas) / len(deltas)
            worst_delta = max(worst_delta, max(deltas))
            total_flips_B += a_flip + fd_flip
            print(f"{did:<10}{budget:>11,}{days:>5}"
                  f"{a_flip:>16}{fd_flip:>16}{mean_delta:>19.1%}")
    print(f"  => recommended-tier flips (accom+food): {total_flips_B} / "
          f"{len(DEST_IDS) * len(SCENARIOS) * N_DRAWS * 2}")
    print(f"  => worst-case max allocation rel-delta under +/-30% weight noise: "
          f"{worst_delta:.1%}")

    print("\n" + "=" * 78)
    print("VERDICT")
    print("  (A) Core decision (which tier is affordable / feasible) is "
          f"{'INVARIANT' if total_flips_A == 0 else 'SENSITIVE'} to weights")
    print("  (B) Advisory allocation tier choices flip in "
          f"{total_flips_B} cases; allocation amounts move <= ~weight noise,")
    print("      bounded by normalization (Sum=1) -> weights are tunable")
    print("      priors, not load-bearing constants.")
    print("=" * 78)


if __name__ == "__main__":
    main()