---
id: wiki-2026-0508-cipomdps
title: CIPOMDPs
category: 10_Wiki/Topics
status: verified
canonical_id: self
aliases: [CI-POMDP, Communicative Interactive POMDP, Cooperative Inverse POMDP]
duplicate_of: none
source_trust_level: A
confidence_score: 0.85
verification_status: applied
tags: [rl, pomdp, multi-agent, alignment]
raw_sources: []
last_reinforced: 2026-05-10
github_commit: pending
tech_stack:
  language: python
  framework: pomdp-py/pettingzoo
---

# CIPOMDPs

## 매 한 줄
> **"매 Cooperative Inverse Partially Observable MDP — 매 human + agent 의 shared reward, 매 reward function 의 hidden parameter."** 매 Hadfield-Menell et al. (CIRL 2016) 의 POMDP extension — 매 assistance games, alignment formalization 의 backbone — 매 2026 에 LLM agent assistance 의 theoretical frame.

## 매 핵심

### 매 정의
- **State** s ∈ S, **Actions** (a^H, a^R) for human + robot.
- **Reward parameter** θ ∈ Θ — 매 human 의 known, robot 의 unknown.
- **Reward** r(s, a^H, a^R; θ) — 매 shared.
- **Observations** o^H, o^R — 매 partial.
- **Goal**: maximize E[Σ r(s, a^H, a^R; θ)] — 매 robot 의 θ 의 inference + acting.

### 매 properties
- **Active learning**: 매 robot 의 information-gathering actions.
- **Off-switch problem**: 매 robot 의 uncertainty 의 corrigibility 의 induce.
- **Reward hacking immunity** (in theory): 매 θ unknown → 매 proxy 의 over-optimize 의 X.

### 매 응용
1. Assistance games (cleaning, cooking robot).
2. RLHF formalization — 매 preference 의 reward 의 evidence.
3. Multi-agent communication (CIPOMDPs with messages).

## 💻 패턴

### CIPOMDP belief update
```python
import numpy as np

def update_theta_belief(b_theta, s, a_H, theta_grid, beta=1.0):
    # Boltzmann human: P(a_H | s, theta) ∝ exp(beta * Q*(s, a_H; theta))
    likelihoods = np.array([
        np.exp(beta * Q_star(s, a_H, theta)) /
        sum(np.exp(beta * Q_star(s, a, theta)) for a in actions_H)
        for theta in theta_grid
    ])
    posterior = b_theta * likelihoods
    return posterior / posterior.sum()
```

### Robot policy (expected utility over θ)
```python
def robot_action(s, b_theta, theta_grid, gamma=0.95):
    # Pick a^R maximizing expected return under belief
    best_a, best_eu = None, -np.inf
    for a_R in actions_R:
        eu = sum(
            b * V_pi(s, a_R, theta, gamma)
            for b, theta in zip(b_theta, theta_grid)
        )
        if eu > best_eu:
            best_a, best_eu = a_R, eu
    return best_a
```

### Off-switch game
```python
def off_switch_decision(b_theta, theta_grid, action_value, switch_off_value=0):
    # Robot defers to human if uncertain about reward
    expected_action_value = sum(
        b * action_value(theta) for b, theta in zip(b_theta, theta_grid)
    )
    # If human can correct, deferring dominates when uncertain
    if expected_action_value < switch_off_value + uncertainty_bonus(b_theta):
        return "wait_for_human"
    return "act"
```

### Active query (info gain)
```python
def best_query(b_theta, theta_grid, candidate_queries):
    def expected_info_gain(q):
        H_prior = entropy(b_theta)
        H_post = sum(
            P_response(r, q, theta) * b *
            entropy(update_theta_belief_query(b_theta, q, r, theta_grid))
            for theta, b in zip(theta_grid, b_theta)
            for r in possible_responses
        )
        return H_prior - H_post
    return max(candidate_queries, key=expected_info_gain)
```

### Boltzmann human model
```python
def boltzmann_human(s, theta, beta=1.0):
    qs = np.array([Q_star(s, a, theta) for a in actions_H])
    probs = np.exp(beta * qs - np.max(beta * qs))
    return probs / probs.sum()
```

## 매 결정 기준
| 상황 | Approach |
|---|---|
| Small θ space | Exact belief update + value iteration |
| Large θ | Particle filter + POMCP/POMCPOW |
| Continuous θ | Variational + amortized inference |
| Real human | Boltzmann + irrationality terms (myopia, bias) |
| Communication | CIPOMDPs with message channel |

**기본값**: 매 particle filter belief + MCTS robot policy.

## 🔗 Graph
- 부모: [[POMDP]]
- 응용: [[RLHF]]
- Adjacent: [[AI Safety and Alignment]] · [[Theory of Mind]]

## 🤖 LLM 활용
**언제**: 매 assistant agent design 의 theoretical justification, 매 ambiguity-handling spec.
**언제 X**: 매 small tactical decision 의 deployment-ready code.

## ❌ 안티패턴
- **Maximize-best-guess θ**: 매 expected utility 의 over Θ — not max-likelihood θ.
- **Rational human assumption**: 매 noisy/biased — 매 Boltzmann + bias models.
- **Static θ**: 매 preferences drift — 매 non-stationary θ 의 model.
- **Ignoring corrigibility**: 매 θ certainty 의 prematurity 의 dangerous.

## 🧪 검증 / 중복
- Verified (Hadfield-Menell et al. NeurIPS 2016, Russell *Human Compatible* 2019).
- 신뢰도 A.

## 🕓 Changelog
| 날짜 | 변경 |
|---|---|
| 2026-05-08 | Phase 1 |
| 2026-05-10 | Manual cleanup — formal CIRL/CIPOMDP with code |