libs/benchmark/policy_repair.py

from __future__ import annotations

from dataclasses import dataclass
from typing import Dict, Iterable, List, Tuple

from libs.adaptive.scheduler import AdaptiveScheduler


@dataclass(frozen=True)
class AdaptivePolicyVariant:
    name: str
    stop_mode: str  # hard | soft | none
    cluster_bootstrap_rounds: int = 0
    min_cluster_coverage: float = 0.0
    min_hypercluster_coverage: float = 0.0
    soft_recovery_batches: int = 0
    soft_threshold_relax: float = 0.0
    soft_model_weight_cap: float = 1.0
    soft_exploration_boost: float = 0.0
    pre_floor_model_weight_cap: float = 1.0
    max_soft_recoveries: int = 1
    epsilon_quality: float = 0.10
    acceptable_regret: float = 0.10
    min_budget: int = 250
    max_budget: int = 10000
    regret_patience_rounds: int = 6
    regret_uncertainty_multiplier: float = 0.5
    regret_min_cluster_coverage: float = 0.20
    regret_min_hypercluster_coverage: float = 0.20


@dataclass
class PolicyRuntimeState:
    soft_recovery_remaining: int = 0
    soft_recoveries_used: int = 0


def default_policy_variants() -> Dict[str, AdaptivePolicyVariant]:
    return {
        "adaptive_hard_stop": AdaptivePolicyVariant(
            name="adaptive_hard_stop",
            stop_mode="hard",
        ),
        "adaptive_soft_stop": AdaptivePolicyVariant(
            name="adaptive_soft_stop",
            stop_mode="soft",
            soft_recovery_batches=16,
            soft_threshold_relax=0.35,
            soft_model_weight_cap=0.45,
            soft_exploration_boost=0.25,
            pre_floor_model_weight_cap=1.0,
            max_soft_recoveries=1,
        ),
        "adaptive_no_stop": AdaptivePolicyVariant(
            name="adaptive_no_stop",
            stop_mode="none",
        ),
        "adaptive_cluster_bootstrap": AdaptivePolicyVariant(
            name="adaptive_cluster_bootstrap",
            stop_mode="hard",
            cluster_bootstrap_rounds=24,
        ),
        "adaptive_soft_stop_cluster_floor": AdaptivePolicyVariant(
            name="adaptive_soft_stop_cluster_floor",
            stop_mode="soft",
            cluster_bootstrap_rounds=20,
            min_cluster_coverage=0.60,
            min_hypercluster_coverage=0.55,
            soft_recovery_batches=20,
            soft_threshold_relax=0.40,
            soft_model_weight_cap=0.40,
            soft_exploration_boost=0.30,
            pre_floor_model_weight_cap=0.35,
            max_soft_recoveries=2,
        ),
    }


def regret_stop_variant(
    *,
    name: str = "adaptive_budgeted_regret_stop",
    epsilon_quality: float = 0.10,
    acceptable_regret: float = 0.10,
    min_budget: int = 250,
    max_budget: int = 10000,
    regret_patience_rounds: int = 6,
    regret_uncertainty_multiplier: float = 0.50,
    regret_min_cluster_coverage: float = 0.20,
    regret_min_hypercluster_coverage: float = 0.20,
) -> AdaptivePolicyVariant:
    return AdaptivePolicyVariant(
        name=name,
        stop_mode="regret",
        epsilon_quality=float(epsilon_quality),
        acceptable_regret=float(acceptable_regret),
        min_budget=int(min_budget),
        max_budget=int(max_budget),
        regret_patience_rounds=int(regret_patience_rounds),
        regret_uncertainty_multiplier=float(regret_uncertainty_multiplier),
        regret_min_cluster_coverage=float(regret_min_cluster_coverage),
        regret_min_hypercluster_coverage=float(regret_min_hypercluster_coverage),
    )


def coverage_fraction(sampled: Iterable[int], total: int) -> float:
    if total <= 0:
        return 0.0
    return float(len(set(sampled)) / float(total))


def coverage_floor_reached(
    variant: AdaptivePolicyVariant,
    *,
    cluster_coverage: float,
    hypercluster_coverage: float,
) -> bool:
    return bool(
        cluster_coverage >= float(variant.min_cluster_coverage)
        and hypercluster_coverage >= float(variant.min_hypercluster_coverage)
    )


def select_cluster_bootstrap_batch(
    scheduler: AdaptiveScheduler,
    *,
    sampled_clusters: set[int],
    batch_size: int,
) -> List[str]:
    active = scheduler.queue_manager.ordered_active()
    if not active:
        return []
    selected: List[str] = []
    # First pass: one candidate from unsampled clusters.
    for item in active:
        if item.cluster_id in sampled_clusters:
            continue
        if item.ligand_id in selected:
            continue
        selected.append(item.ligand_id)
        if len(selected) >= batch_size:
            break
    # Fill remaining by priority order.
    if len(selected) < batch_size:
        for item in active:
            if item.ligand_id in selected:
                continue
            selected.append(item.ligand_id)
            if len(selected) >= batch_size:
                break
    for ligand_id in selected:
        scheduler.queue_manager.items[ligand_id].times_selected += 1
    return selected


def effective_controls(
    variant: AdaptivePolicyVariant,
    state: PolicyRuntimeState,
    *,
    coverage_floor_active: bool,
) -> Tuple[float, float, float]:
    # Returns (model_weight_cap, exploration_boost, threshold_relax).
    model_cap = 1.0
    exploration_boost = 0.0
    threshold_relax = 0.0
    if coverage_floor_active:
        model_cap = min(model_cap, float(variant.pre_floor_model_weight_cap))
    if state.soft_recovery_remaining > 0:
        model_cap = min(model_cap, float(variant.soft_model_weight_cap))
        exploration_boost = max(exploration_boost, float(variant.soft_exploration_boost))
        threshold_relax = max(threshold_relax, float(variant.soft_threshold_relax))
    return float(model_cap), float(exploration_boost), float(threshold_relax)


def apply_stop_policy(
    variant: AdaptivePolicyVariant,
    state: PolicyRuntimeState,
    *,
    stop_signal: bool,
    stop_reason: str,
    coverage_floor_active: bool,
) -> Tuple[bool, str, PolicyRuntimeState]:
    mode = str(variant.stop_mode).strip().lower()
    # Coverage floor blocks stop decisions in floor-enforced variants.
    if coverage_floor_active and stop_signal:
        return False, "coverage_floor_blocked_stop", state

    if mode == "none":
        return False, "no_stop_policy", state

    if mode == "hard":
        return bool(stop_signal), (stop_reason if stop_signal else "continue"), state

    if mode == "regret":
        return bool(stop_signal), (stop_reason if stop_signal else "continue"), state

    # soft mode
    if state.soft_recovery_remaining > 0:
        state.soft_recovery_remaining -= 1
        if state.soft_recovery_remaining <= 0 and stop_signal:
            if state.soft_recoveries_used >= int(variant.max_soft_recoveries):
                return True, "soft_recovery_exhausted", state
            state.soft_recovery_remaining = int(variant.soft_recovery_batches)
            state.soft_recoveries_used += 1
            return False, "soft_recovery_restarted", state
        return False, "soft_recovery_active", state

    if not stop_signal:
        return False, "continue", state

    if state.soft_recoveries_used >= int(variant.max_soft_recoveries):
        return True, "soft_recovery_limit_reached", state

    state.soft_recovery_remaining = int(variant.soft_recovery_batches)
    state.soft_recoveries_used += 1
    return False, "soft_recovery_started", state