File size: 14,578 Bytes
654c8c7 6ad7bd8 654c8c7 b5e5650 52a986a 654c8c7 6ad7bd8 7ccb648 f55f75f 6ad7bd8 52a986a 6ad7bd8 654c8c7 4b5c463 654c8c7 b5e5650 654c8c7 5144b7e 654c8c7 52a986a 654c8c7 52a986a 654c8c7 52a986a 654c8c7 52a986a 654c8c7 52a986a 654c8c7 52a986a 654c8c7 52a986a 654c8c7 52a986a 6ad7bd8 | 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 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 286 287 288 289 290 291 292 293 294 295 296 297 298 299 300 301 302 303 304 305 306 307 308 309 310 311 312 313 314 315 316 317 318 319 320 321 322 323 324 325 326 327 328 329 330 331 332 333 334 335 336 337 338 339 340 341 342 343 344 345 346 347 348 349 350 351 352 353 354 355 356 357 358 359 360 361 362 363 364 365 366 367 368 369 370 371 372 373 374 375 376 377 378 379 380 381 382 383 384 385 386 387 388 389 390 391 392 393 394 395 396 397 398 399 400 401 402 403 404 405 | """
AntiAtropos Stability Layer — Phase 3.
This module is the mathematical core of the Lyapunov-inspired reward signal.
It is intentionally a stateless collection of pure functions so it can be
swapped, extended, or unit-tested independently of the environment.
Key concepts implemented
------------------------
1. Lyapunov Energy V(s) = Σ Q_i²
The "potential energy" of the cluster. Zero means all queues are empty.
A rising V means the cluster is destabilising.
2. Lyapunov Drift ΔV(t) = V(s_t) − V(s_{t−1})
The primary stabilising objective. The reward penalises positive drift.
Negative drift (energy decreasing) is "good" — the agent drove the system
toward the equilibrium.
3. Control-Barrier Function h_i(s) = max(0, Q_i − Q_max)²
A soft safety constraint. h_i > 0 only when node i has exceeded the
hard-queue-depth safety ceiling Q_max. Summing over all nodes gives the
total barrier violation penalty.
4. Neely Drift-Plus-Penalty (optional, advanced)
From Neely's Lyapunov optimisation framework:
Δ(t) + V_weight · p(t)
where:
Δ(t) = ΔV — the one-step Lyapunov drift.
p(t) = cost — the per-step penalty (infrastructure cost here).
V_weight — trade-off parameter: large V_weight → prioritise cost;
small V_weight → prioritise stability.
Minimising this expression at every step produces a policy that is both
stable (queue-stable in the mean) and cost-efficient.
"""
from __future__ import annotations
import os
import math
import statistics
from typing import Sequence
# ---------------------------------------------------------------------------
# Safety ceiling used by the barrier function
# ---------------------------------------------------------------------------
Q_BARRIER_MAX: float = 150.0
"""Queue depth above which the barrier function fires (hard safety zone).
Set higher than OVERLOAD_THRESHOLD (80) to allow the agent time to react
before the barrier penalty kicks in."""
MAX_QUEUE_NORM: float = 200.0
"""Normalization divisor shared with environment.py. Matches FATAL_FAIL_THRESHOLD."""
BARRIER_NORM_SCALE: float = 10000.0
"""Normalization divisor for the barrier term.
The raw barrier H(s) = sum(max(0, Q_i - Q_max)^2) can produce very large values
(e.g. 5 nodes at Q=200, Q_max=150 gives 5*2500=12500). Without normalization,
this dominates the reward. Dividing by this scale keeps barrier in the same
order of magnitude as the other terms when delta=0.005."""
STABILITY_WINDOW: int = 10
"""Number of ticks to look back when judging whether the system is
trend-stable (V is on a decreasing trajectory)."""
# ---------------------------------------------------------------------------
# Reward normalisation defaults (env-overridable)
# ---------------------------------------------------------------------------
REWARD_NORM_MIDPOINT: float = float(os.getenv("ANTIATROPOS_REWARD_MIDPOINT", "-6.0"))
REWARD_NORM_TEMPERATURE: float = float(os.getenv("ANTIATROPOS_REWARD_TEMPERATURE", "2.0"))
REWARD_NORM_EPS: float = float(os.getenv("ANTIATROPOS_REWARD_EPS", "1e-8"))
REWARD_SCALE_VERSION: str = "sigmoid-v2" # v2: smooth SLA + barrier active
# ---------------------------------------------------------------------------
# Core Lyapunov functions
# ---------------------------------------------------------------------------
def compute_lyapunov(nodes: list[dict]) -> float:
"""
V(s) = Σ Q_i²
Sum of squared queue depths across all nodes. This is the cluster's
Lyapunov energy. Lower is more stable; zero means all queues are empty.
Args:
nodes: List of per-node state dicts (as returned by simulator.state()).
Each dict must contain the key ``queue_depth``.
Returns:
Scalar Lyapunov energy ≥ 0.
"""
return float(
sum(
float(n.get("importance_weight", 1.0)) * (n["queue_depth"] ** 2)
for n in nodes
)
)
def compute_lyapunov_graph(
nodes: list[dict],
topology: dict[str, list[str]],
edge_weight: float = 5.0,
) -> float:
"""
V_graph(s) = Σ w_i·Q_i² + edge_weight · Σ_{(i,j)∈edges} |Q_i - Q_j|
The edge term penalises flow imbalance between connected nodes.
If node-0 is overloaded but node-1 (its child) is idle, the edge
term fires even if node-1's individual energy contribution is zero.
This gives the agent gradient signal to balance load across the graph,
not just reduce individual node queues.
With edge_weight=5.0 and max |Q_diff|=200 on 4 edges, the edge term
adds ~4000 to the Lyapunov energy, or ~2-5 % of the base energy at
full overload — a meaningful secondary gradient without dominating.
"""
node_map = {n["node_id"]: n for n in nodes}
# Standard weighted Lyapunov
base_energy = compute_lyapunov(nodes)
# Edge imbalance penalty (raw queue-depth differences)
edge_penalty = 0.0
for parent_id, children in topology.items():
parent = node_map.get(parent_id)
if not parent:
continue
for child_id in children:
child = node_map.get(child_id)
if not child:
continue
imbalance = abs(float(parent["queue_depth"]) - float(child["queue_depth"]))
edge_penalty += imbalance
return base_energy + edge_weight * edge_penalty
def compute_drift(v_prev: float, v_curr: float) -> float:
"""
ΔV(t) = V(s_t) − V(s_{t−1})
One-step drift in Lyapunov energy.
Negative drift → the agent moved the system toward a lower-energy state.
Positive drift → the cluster is destabilising.
Args:
v_prev: Lyapunov energy at the *previous* tick.
v_curr: Lyapunov energy at the *current* tick.
Returns:
Signed scalar drift value.
"""
return v_curr - v_prev
def compute_barrier(nodes: list[dict], q_max: float = Q_BARRIER_MAX) -> float:
"""
Control-Barrier Function (CBF) violation penalty.
H(s) = Σ_i max(0, Q_i − Q_max)²
This is zero when no node exceeds the ceiling and grows quadratically as
queues enter the "hard danger zone" above Q_max. It can be added to the
reward as an extra penalty for unsafe states.
Args:
nodes: Per-node state dicts (must contain ``queue_depth``).
q_max: Safety ceiling for queue depth. Default: Q_BARRIER_MAX.
Returns:
Scalar barrier violation energy ≥ 0.
"""
violation = 0.0
for n in nodes:
excess = n["queue_depth"] - q_max
if excess > 0:
violation += excess ** 2
return violation
# ---------------------------------------------------------------------------
# Trend analysis
# ---------------------------------------------------------------------------
def is_lyapunov_stable(
v_history: Sequence[float],
window: int = STABILITY_WINDOW,
) -> bool:
"""
Return True if the Lyapunov energy has been on a non-increasing trend
over the last ``window`` ticks.
Uses a simple linear regression slope: if slope ≤ 0 the system is
considered trend-stable.
Args:
v_history: Ordered sequence of Lyapunov energy values (oldest first).
window: How many recent values to consider.
Returns:
True if the system is trend-stable, False otherwise.
"""
recent = list(v_history[-window:])
if len(recent) < 2:
return True # not enough data — assume stable at episode start
n = len(recent)
xs = list(range(n))
mean_x = (n - 1) / 2.0
mean_y = statistics.mean(recent)
num = sum((xs[i] - mean_x) * (recent[i] - mean_y) for i in range(n))
den = sum((xs[i] - mean_x) ** 2 for i in range(n))
if den == 0:
return True
slope = num / den
return slope <= 0.0
def lyapunov_variance(v_history: Sequence[float]) -> float:
"""
Variance of the Lyapunov energy trajectory over an episode.
Used by the grader as the primary stability metric: a lower variance
means the agent kept the cluster in a consistently stable state, rather
than allowing wild oscillations.
Args:
v_history: All per-tick V(s) values for the episode.
Returns:
Population variance of the energy trajectory.
"""
if len(v_history) < 2:
return 0.0
return statistics.variance(v_history)
# ---------------------------------------------------------------------------
# Neely Drift-Plus-Penalty (advanced reward signal)
# ---------------------------------------------------------------------------
def drift_plus_penalty(
v_prev: float,
v_curr: float,
penalty_cost: float,
V_weight: float = 1.0,
) -> float:
"""
Neely's Drift-Plus-Penalty objective:
DPP(t) = ΔV(t) + V_weight · p(t)
where:
ΔV(t) = v_curr − v_prev (Lyapunov drift)
p(t) = penalty_cost (infrastructure cost this tick)
V_weight = trade-off coefficient:
large → agent optimises cost more aggressively,
small → agent focuses on stability.
Minimising this at each step produces a queue-stable policy with bounded
average cost — the theoretical guarantee from Neely's framework.
This function can substitute for the simpler ΔV term in the reward
when you want to make the cost trade-off explicit and theoretically
grounded (rather than the ad-hoc β·Cost term).
Args:
v_prev: Lyapunov energy at previous tick.
v_curr: Lyapunov energy at current tick.
penalty_cost: Per-step cost to penalise (e.g. current_cost_per_hour).
V_weight: Trade-off weight V in Neely's framework.
Returns:
Scalar DPP value. The reward should negate this:
R_t = −DPP(t) − γ·SLA_violation_step
"""
delta_v = compute_drift(v_prev, v_curr)
return delta_v + V_weight * penalty_cost
# ---------------------------------------------------------------------------
# Convenience: full reward computation (matches environment.py formula)
# ---------------------------------------------------------------------------
def smooth_sla_penalty(
avg_latency_norm: float,
error_rate: float,
latency_threshold: float = 0.20,
error_threshold: float = 0.05,
latency_temperature: float = 0.03,
error_temperature: float = 0.01,
) -> float:
"""
Smooth SLA penalty in [0, 1] that ramps up as metrics approach thresholds.
Unlike the binary cliff (0 or 1), this gives the agent gradient signal
BEFORE the SLA is actually violated, enabling preventive learning.
Uses two sigmoids (one for latency, one for errors) and takes the max
so whichever dimension is worse dominates.
Args:
avg_latency_norm: Normalized average latency [0, 1].
error_rate: Cluster-wide error rate [0, 1].
latency_threshold: Normalized latency SLA boundary.
error_threshold: Error rate SLA boundary.
latency_temperature: Sigmoid temperature for latency (lower = sharper).
error_temperature: Sigmoid temperature for errors (lower = sharper).
Returns:
Smooth penalty in [0, 1]. Near 0 when safe, near 1 when violating.
Raises:
ValueError: If inputs are outside [0, 1], indicating raw (non-normalized)
values were passed by mistake. This is a common bug: passing latency
in raw ms (e.g. 200.0) instead of normalized [0,1] (e.g. 0.20).
"""
if avg_latency_norm < -0.01 or avg_latency_norm > 1.5:
raise ValueError(
f"smooth_sla_penalty: avg_latency_norm={avg_latency_norm:.4f} is outside "
f"expected [0, 1] range. Did you pass raw ms instead of normalized? "
f"Divide by MAX_LATENCY_NORM before calling."
)
if error_rate < -0.01 or error_rate > 1.5:
raise ValueError(
f"smooth_sla_penalty: error_rate={error_rate:.4f} is outside "
f"expected [0, 1] range."
)
lat_z = (avg_latency_norm - latency_threshold) / max(1e-8, latency_temperature)
err_z = (error_rate - error_threshold) / max(1e-8, error_temperature)
lat_penalty = 1.0 / (1.0 + math.exp(-lat_z))
err_penalty = 1.0 / (1.0 + math.exp(-err_z))
return max(lat_penalty, err_penalty)
def compute_reward(
v_prev: float,
v_curr: float,
cost: float,
sla_violation_step: float = 0.0,
alpha: float = 1.0,
beta: float = 0.05,
gamma: float = 2.0,
barrier: float = 0.0,
delta: float = 0.005,
) -> float:
"""
R_t = -(alpha * DeltaV(s) + beta * Cost + gamma * SLA_smooth + delta * Barrier)
Convenience wrapper that mirrors the reward formula in environment.py.
Args:
v_prev: Lyapunov energy at previous tick.
v_curr: Lyapunov energy at current tick.
cost: Infrastructure cost this tick (USD/hr).
sla_violation_step: Smooth SLA penalty in [0, 1] (was binary 0/1).
alpha: Weight on Lyapunov drift.
beta: Weight on cost.
gamma: Weight on SLA violations.
barrier: Control-barrier function violation energy.
delta: Weight on barrier penalty.
Returns:
Scalar reward (higher is better, always <= 0 in a stable episode).
"""
delta_v = compute_drift(v_prev, v_curr)
# Normalize barrier to prevent reward domination: raw barrier can be ~12500,
# after dividing by BARRIER_NORM_SCALE it's ~1.25, then scaled by delta=0.005
# gives ~0.006 which is comparable to other terms.
barrier_normalized = barrier / BARRIER_NORM_SCALE if BARRIER_NORM_SCALE > 0 else barrier
return -(alpha * delta_v + beta * cost + gamma * sla_violation_step + delta * barrier_normalized)
def normalize_reward(
raw_reward: float,
midpoint: float = REWARD_NORM_MIDPOINT,
temperature: float = REWARD_NORM_TEMPERATURE,
eps: float = REWARD_NORM_EPS,
) -> float:
"""
Deterministically map raw reward to [0, 1] using a smooth sigmoid.
reward_01 = 1 / (1 + exp(-(raw_reward - midpoint) / temperature))
"""
temp = max(float(eps), abs(float(temperature)))
z = (float(raw_reward) - float(midpoint)) / temp
if z >= 0:
exp_neg = math.exp(-z)
out = 1.0 / (1.0 + exp_neg)
else:
exp_pos = math.exp(z)
out = exp_pos / (1.0 + exp_pos)
return min(1.0, max(0.0, float(out)))
|