Spaces:
Running
Running
File size: 8,839 Bytes
df97e68 | 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 | """
Converts ObservationModel (Pydantic) → flat numpy float32 vector.
All downstream RL code depends on OBS_DIM being stable.
Feature layout (total = OBS_DIM = 84):
[0 : 63) — per-service block (7 services × 9 features each)
[63 : 84) — global block (21 scalar features)
"""
from __future__ import annotations
import numpy as np
from typing import List
from app.models import (
ObservationModel,
ServiceType,
StageType,
PriorityMode,
ActionType,
)
# ── Canonical orderings (must never change across the codebase) ──────────────
SERVICES: List[ServiceType] = [
ServiceType.PASSPORT,
ServiceType.DRIVING_LICENSE,
ServiceType.GST_REGISTRATION,
ServiceType.INCOME_CERTIFICATE,
ServiceType.CASTE_CERTIFICATE,
ServiceType.BIRTH_CERTIFICATE,
ServiceType.LAND_REGISTRATION,
] # 7 legacy RL services for checkpoint compatibility
STAGES: List[StageType] = list(StageType) # 5 stages
PRIORITY_MODES: List[PriorityMode] = list(PriorityMode) # 4 modes
ACTION_TYPES: List[ActionType] = list(ActionType) # 6 types
SERVICE_IDX = {s: i for i, s in enumerate(SERVICES)}
STAGE_IDX = {s: i for i, s in enumerate(STAGES)}
PM_IDX = {m: i for i, m in enumerate(PRIORITY_MODES)}
AT_IDX = {a: i for i, a in enumerate(ACTION_TYPES)}
# ── Dimension constants ───────────────────────────────────────────────────────
N_SERVICES = len(SERVICES) # 7
N_STAGES = len(STAGES) # 5
N_PRIORITY_MODES = len(PRIORITY_MODES) # 4
N_ACTION_TYPES = len(ACTION_TYPES) # 6
PER_SERVICE_DIM = 4 + N_STAGES # queue_len, avg_wait, urgent, missing + 5 stage fracs = 9
GLOBAL_DIM = (
1 # day_ratio
+ 1 # total_backlog_normalized
+ 1 # total_completed_normalized
+ 1 # total_sla_breaches_normalized
+ 1 # fairness_gap
+ 1 # escalation_budget_ratio
+ 1 # last_action_valid
+ N_ACTION_TYPES # last_action_type one-hot (6)
+ N_PRIORITY_MODES # current_priority_mode one-hot (4)
+ 1 # idle_officer_ratio
+ 1 # urgent_backlog_ratio
+ 1 # officer_utilization
+ 1 # backlog_per_officer
) # = 21
OBS_DIM = N_SERVICES * PER_SERVICE_DIM + GLOBAL_DIM # 63 + 21 = 84
# ── Normalisation caps (avoid div-by-zero, keep values in [0,1]) ─────────────
_MAX_QUEUE = 200.0
_MAX_WAIT = 30.0
_MAX_URGENT = 50.0
_MAX_MISSING = 50.0
_MAX_BACKLOG = 500.0
_MAX_COMPLETED = 500.0
_MAX_SLA = 100.0
_MAX_ESC_BUDGET = 20.0
_MAX_OFFICERS = 50.0
class FeatureBuilder:
"""
Stateless transformer: ObservationModel → np.ndarray[float32, OBS_DIM].
Usage:
fb = FeatureBuilder()
vec = fb.build(obs, current_priority_mode="urgent_first",
last_action_type="advance_time")
"""
def build(
self,
obs: ObservationModel,
current_priority_mode: str = "balanced",
last_action_type: str = "advance_time",
) -> np.ndarray:
features = np.zeros(OBS_DIM, dtype=np.float32)
offset = 0
snap_dict = {
snap.service_type: snap
for snap in (obs.queue_snapshots.values() if isinstance(obs.queue_snapshots, dict) else obs.queue_snapshots)
}
# ── Per-service block ─────────────────────────────────────────────
for svc in SERVICES:
snap = snap_dict.get(svc)
if snap is None:
offset += PER_SERVICE_DIM
continue
total_in_svc = max(getattr(snap, "total_pending", getattr(snap, "active_cases", 0)), 1)
features[offset + 0] = getattr(snap, "total_pending", getattr(snap, "active_cases", 0)) / _MAX_QUEUE
features[offset + 1] = getattr(snap, "avg_age_days", getattr(snap, "avg_waiting_days", getattr(snap, "oldest_case_age_days", 0))) / _MAX_WAIT
features[offset + 2] = getattr(snap, "urgent_cases", getattr(snap, "urgent_pending", 0)) / _MAX_URGENT
features[offset + 3] = getattr(snap, "blocked_missing_docs", getattr(snap, "missing_docs_cases", 0)) / _MAX_MISSING
# Stage distribution as fractions
stage_counts = getattr(snap, "stage_counts", getattr(snap, "public_stage_counts", {})) or {}
for stg in STAGES:
count = stage_counts.get(stg, 0)
features[offset + 4 + STAGE_IDX[stg]] = count / total_in_svc
offset += PER_SERVICE_DIM
# ── Global block ──────────────────────────────────────────────────
day_ratio = obs.day / max(obs.max_days, 1)
features[offset + 0] = day_ratio
features[offset + 1] = obs.total_backlog / _MAX_BACKLOG
features[offset + 2] = obs.total_completed / _MAX_COMPLETED
features[offset + 3] = obs.total_sla_breaches / _MAX_SLA
features[offset + 4] = float(getattr(obs, "fairness_gap", getattr(obs, "fairness_index", 0.0)) or 0.0)
features[offset + 5] = obs.escalation_budget_remaining / _MAX_ESC_BUDGET
features[offset + 6] = float(obs.last_action_valid)
offset += 7
# Last action type one-hot
at_vec = np.zeros(N_ACTION_TYPES, dtype=np.float32)
try:
at_vec[AT_IDX[ActionType(last_action_type)]] = 1.0
except (ValueError, KeyError):
pass
features[offset: offset + N_ACTION_TYPES] = at_vec
offset += N_ACTION_TYPES
# Current priority mode one-hot
pm_vec = np.zeros(N_PRIORITY_MODES, dtype=np.float32)
try:
pm_vec[PM_IDX[PriorityMode(current_priority_mode)]] = 1.0
except (ValueError, KeyError):
pass
features[offset: offset + N_PRIORITY_MODES] = pm_vec
offset += N_PRIORITY_MODES
# Officer-derived scalars
pool = obs.officer_pool
total_officers = max(getattr(pool, "total_officers", 1) if not callable(getattr(pool, "total_officers", None)) else pool.total_officers(), 1)
idle_officers = getattr(pool, "idle_officers", getattr(pool, "reserve_officers", 0))
idle_ratio = idle_officers / total_officers
total_backlog_safe = max(obs.total_backlog, 1)
urgent_total = sum(
getattr(snap_dict[s], "urgent_cases", getattr(snap_dict[s], "urgent_pending", 0))
for s in SERVICES
if s in snap_dict
)
urgent_ratio = urgent_total / total_backlog_safe
utilization = (total_officers - idle_officers) / total_officers
backlog_per_off = obs.total_backlog / total_officers
features[offset + 0] = float(np.clip(idle_ratio, 0.0, 1.0))
features[offset + 1] = float(np.clip(urgent_ratio, 0.0, 1.0))
features[offset + 2] = float(np.clip(utilization, 0.0, 1.0))
features[offset + 3] = float(np.clip(backlog_per_off / _MAX_OFFICERS, 0.0, 1.0))
assert offset + 4 == OBS_DIM, f"OBS_DIM mismatch: {offset + 4} != {OBS_DIM}"
return features
# -- Action space layout (N_ACTIONS = 28) -------------------------------------
#
# 0 - 3 : set_priority_mode (4 modes in PRIORITY_MODES order)
# 4 - 10 : request_missing_documents per service (7)
# 11 - 17 : escalate_service per service (7)
# 18 : advance_time
# 19 - 25 : reallocate_officers from source service -> most loaded other service
# 26 : assign_capacity +1 to most-loaded service
# 27 : assign_capacity +1 to most-urgent service
N_ACTIONS = 4 + N_SERVICES + N_SERVICES + 1 + N_SERVICES + 2 # = 28
ACTION_DECODE_TABLE = {}
idx = 0
for m in PRIORITY_MODES:
ACTION_DECODE_TABLE[idx] = ("set_priority_mode", None, m.value, None)
idx += 1
for s in SERVICES:
ACTION_DECODE_TABLE[idx] = ("request_missing_documents", s.value, None, None)
idx += 1
for s in SERVICES:
ACTION_DECODE_TABLE[idx] = ("escalate_service", s.value, None, None)
idx += 1
ACTION_DECODE_TABLE[idx] = ("advance_time", None, None, None); idx += 1
for s in SERVICES:
ACTION_DECODE_TABLE[idx] = ("reallocate_officers", s.value, "most_loaded_other", 1)
idx += 1
ACTION_DECODE_TABLE[idx] = ("assign_capacity", "__most_loaded__", None, 1); idx += 1
ACTION_DECODE_TABLE[idx] = ("assign_capacity", "__most_urgent__", None, 1); idx += 1
assert len(ACTION_DECODE_TABLE) == N_ACTIONS
|