Files

T

Antigravity Agent f8b21af4be Wiki cleanup: error-doc removal, dedup merge, link normalization

10_Wiki/Topics 대규모 정리:
- 오류 캡처/미완성 stub 문서 227개 제거
- 교차폴더 중복 43클러스터 병합 (63파일 → redirect)
- 링크명 정규화: 깨진 링크 수정·redirect 직결·개념 매핑 ~2,400건
- 카테고리 MOC 6개 신규 생성
- Graph 섹션 미해결 related-keyword 링크 10,058건 제거

Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>

2026-05-20 23:52:15 +09:00

9.0 KiB

Raw Blame History

id, title, category, status, canonical_id, aliases, duplicate_of, source_trust_level, confidence_score, verification_status, tags, raw_sources, last_reinforced, github_commit, tech_stack

title

Cybernetics

매 한 줄

"매 living + 매 machine 의 control + communication 의 universal". Norbert Wiener (1948). 매 feedback loop + homeostasis + 매 information transmission. 매 AI / robotics / biology / sociology 의 cross-disciplinary roots. 매 modern: 매 multi-agent system + 매 control + 매 active inference.

매 핵심

Wiener (1st-order cybernetics)

매 system 의 observe.
매 input → process → output → feedback.
매 negative feedback 의 stabilize.
매 positive feedback 의 amplify / explode.

2nd-order cybernetics (Foerster, Maturana)

매 observer 의 part of system.
매 self-reference.
매 autopoiesis (self-creating).

매 핵심 concept

Feedback Loop

Negative: 매 thermostat, 매 homeostasis.
Positive: 매 microphone feedback, 매 viral growth.

Homeostasis

매 internal state 의 stable.
매 perturbation 의 counter.
매 biological + 매 artificial.

Black Box

매 internal X — 매 I/O 만.
매 abstraction principle.

Variety (Ashby's Law)

매 controller 의 variety ≥ disturbance 의 variety.
매 system 의 control 의 limit.

Requisite Variety

매 system 의 manage 의 environment 의 capable variety.

매 history

1948 Wiener "Cybernetics".
1956 Dartmouth (AI 이름).
Macy Conferences (1946-1953): cyber + systems.
2nd-order: Heinz von Foerster, Maturana-Varela.
Decline (1970s): AI vs Cybernetics 의 split.
Revival (2000s+): control + agent + autopoiesis.

매 modern relevance

Control theory

PID, MPC, optimal control.
Robotics: 매 closed-loop.
매 autonomous vehicle.

Multi-agent system

매 distributed control.
매 swarm.

Active inference (Friston)

매 free energy minimization.
매 cybernetic + Bayesian.
Bayesian-Brain-Hypothesis 참조.

Organizational

매 systems thinking (Senge "Fifth Discipline").
매 viable system model (Beer).

AI alignment

매 reward hacking 의 cybernetic 의 lens.
매 unintended feedback.

매 응용

Thermostat / HVAC: 매 simplest.
Robotics: 매 control.
Biology: 매 endocrine / neural.
Economics: 매 supply-demand.
Organization: 매 managed system.
AI agent: 매 action + observation loop.

💻 패턴

PID Controller (classic feedback)

class PID:
    def __init__(self, kp=1.0, ki=0.1, kd=0.05):
        self.kp, self.ki, self.kd = kp, ki, kd
        self.integral = 0
        self.prev_error = 0
    
    def update(self, target, current, dt):
        error = target - current
        self.integral += error * dt
        derivative = (error - self.prev_error) / dt
        
        output = self.kp * error + self.ki * self.integral + self.kd * derivative
        self.prev_error = error
        return output

# 매 example: 매 thermostat
pid = PID()
while True:
    error_correction = pid.update(target_temp=22, current=sensor.read(), dt=1.0)
    heater.set_power(error_correction)
    sleep(1)

Negative feedback (homeostasis)

class Homeostat:
    def __init__(self, target):
        self.target = target
        self.state = target
    
    def perturbate(self, disturbance):
        self.state += disturbance
    
    def regulate(self, gain=0.1):
        # 매 negative feedback 의 self-correct
        error = self.target - self.state
        self.state += gain * error

# 매 simulate
h = Homeostat(target=37)  # 매 body temp
for _ in range(100):
    h.perturbate(random.gauss(0, 1))  # 매 environment
    h.regulate()
print(h.state)  # 매 ~37

Positive feedback (caution)

def viral_growth(initial, gain, steps):
    """매 positive feedback — 매 amplify."""
    state = initial
    for _ in range(steps):
        state *= (1 + gain)
        if state > LIMIT: break  # 매 saturation needed
    return state

Ashby's Law check

def can_control(controller_variety, disturbance_variety):
    """매 'Only variety can destroy variety.'"""
    return controller_variety >= disturbance_variety

# 매 example: 매 thermostat 의 매 1 mode 만 → 매 매 disturbance 매 multiple 의 fail
print(can_control(1, 5))  # False
print(can_control(10, 5))  # True

Black Box (system identification)

import numpy as np

def identify_black_box(system_fn, n_samples=100):
    """매 input-output 의 model 의 fit."""
    X = np.random.randn(n_samples, INPUT_DIM)
    Y = np.array([system_fn(x) for x in X])
    
    from sklearn.linear_model import LinearRegression
    model = LinearRegression().fit(X, Y)
    return model  # 매 estimated transfer function

Active inference (Friston-style)

def active_inference(belief, world_model, possible_actions):
    """매 minimize expected free energy."""
    best_action, best_efe = None, float('inf')
    for a in possible_actions:
        next_belief = world_model.predict(belief, a)
        # 매 epistemic value
        info_gain = expected_kl(next_belief, belief)
        # 매 pragmatic value
        pragmatic = expected_log_preference(next_belief)
        efe = -info_gain - pragmatic
        if efe < best_efe:
            best_efe, best_action = efe, a
    return best_action

MPC (Model Predictive Control)

import cvxpy as cp

def mpc_step(x_current, x_target, horizon=10):
    x = cp.Variable((horizon + 1, STATE_DIM))
    u = cp.Variable((horizon, ACTION_DIM))
    
    cost = 0
    constraints = [x[0] == x_current]
    for t in range(horizon):
        cost += cp.sum_squares(x[t+1] - x_target) + 0.1 * cp.sum_squares(u[t])
        constraints += [x[t+1] == A @ x[t] + B @ u[t]]
        constraints += [cp.abs(u[t]) <= U_MAX]
    
    cp.Problem(cp.Minimize(cost), constraints).solve()
    return u[0].value

Multi-agent feedback (consensus)

def consensus_step(agents, alpha=0.1):
    """매 distributed averaging."""
    new_states = []
    for a in agents:
        avg_neighbor = mean(n.state for n in a.neighbors)
        new_states.append(a.state + alpha * (avg_neighbor - a.state))
    for a, ns in zip(agents, new_states):
        a.state = ns

# 매 매 step 의 모두 의 average 의 converge.

Reward hacking detection (cybernetic perspective)

def detect_reward_hacking(agent_trajectory, true_objective):
    """매 unintended feedback 의 detect."""
    declared_reward = sum(t.reward for t in agent_trajectory)
    actual_progress = measure_against_intent(agent_trajectory, true_objective)
    
    if declared_reward > 0.9 * MAX_DECLARED and actual_progress < 0.5:
        return 'WARN: reward hacking — 매 metric 의 game'
    return 'OK'

🤔 결정 기준

응용	Approach
Setpoint regulation	PID
Multi-state control	MPC
Distributed	Consensus protocol
Homeostatic system	Negative feedback
Growth (controlled)	Positive feedback + saturation
Multi-agent	Active inference / consensus
AI alignment	Reward hacking detection

기본값: PID 의 baseline. 매 model-aware = MPC. 매 multi-agent = active inference.

🔗 Graph

부모: Systems-Theory · Control-Theory · Multi-agent-System
변형: Feedback-Loop · Homeostasis (항상성) · Autopoiesis
사상가: Norbert-Wiener
응용: PID · MPC · Active-Inference · Bayesian-Brain-Hypothesis
Adjacent: Antifragility · Artificial-Life · Biological-Intelligence · Anarchism

🤖 LLM 활용

언제: 매 control system. 매 multi-agent. 매 active inference. 매 cybernetic / systems thinking. 언제 X: 매 specific math (control theory 의 textbook). 매 single-step decision.

❌ 안티패턴

Open-loop control: 매 perturbation 의 ignore.
Variety mismatch (Ashby): 매 controller 의 weak.
Positive feedback 의 unbounded: 매 explosion.
Reward hacking 의 ignore: 매 unintended.
2nd-order observer 의 X: 매 self-reference 의 limit.

🧪 검증 / 중복

Verified (Wiener "Cybernetics", Ashby "Introduction to Cybernetics", Friston FEP).
신뢰도 A.
Related: Bayesian-Brain-Hypothesis · Antifragility · Multi-agent-System · Anarchism · Computational-Neuroscience-RL.

🕓 Changelog

날짜	변경
2026-05-08	Phase 1
2026-05-10	Manual cleanup — Wiener + 2nd-order + Ashby + 매 PID / MPC / consensus / active inference code

9.0 KiB Raw Blame History