SLCD-v0

Category: Validated Case Study Agents: 2 (Samsung, Sony) Difficulty: Intermediate Source: coopetition_gym/envs/case_study_envs.py

Overview

SLCD Phases

SLCD-v0 implements the Samsung-Sony S-LCD Joint Venture as described in TR-1 (arXiv:2510.18802). This is the gold standard benchmark for Coopetition-Gym, with parameters validated against real business data.

The environment models the 2004-2012 joint venture where Samsung and Sony collaborated on LCD panel manufacturing while competing in the consumer electronics market, a quintessential coopetitive relationship.

Validation Score: 58/60 against historical data

MARL Classification

Property Value
Game Type Markov Game (2-player, general-sum) with empirically validated parameters
Cooperation Structure Mixed-Motive coopetition (cooperation on panels, competition on TVs)
Observability Full
Communication Implicit (through actions only)
Agent Symmetry Near-symmetric (Samsung slight technology advantage: α=0.55 vs 0.45)
Reward Structure Mixed with validated interdependence
Action Space Continuous: A_i = [0, 100]
State Dynamics Deterministic
Horizon Finite, T = 100
Canonical Comparison Empirically calibrated coopetition; cf. Ritala & Hurmelinna-Laukkanen (2009)

Validation Status: Parameters derived from TR-1 §8.3, achieving 58/60 accuracy against historical S-LCD data (2004-2012).

Formal Specification

This environment is formalized as a 2-player Markov Game with empirically validated parameters from the Samsung-Sony S-LCD joint venture.

Agents

N = {Samsung, Sony}

Agent Index Endowment Baseline Bargaining α Role
Samsung 0 100.0 30.0 0.55 Technology provider
Sony 1 100.0 30.0 0.45 Market/brand provider

Samsung’s slight bargaining advantage (55% vs 45%) reflects technology control.

State Space

S ⊆ ℝ¹⁷ (standard dyadic structure)

Action Space

A_i = [0, 100] ⊂ ℝ representing investment in joint venture operations.

Uniaxial Treatment: This environment uses the single-dimension action space characteristic of Coopetition-Gym v1.x. Samsung-Sony competition in consumer TVs is captured through bargaining shares (α=0.55/0.45) rather than explicit competitive actions.

Validated Trust Parameters (TR-1 §8.3)

Parameter Symbol Value Validation Source
Trust Building Rate λ⁺ 0.08 Historical collaboration phases
Trust Erosion Rate λ⁻ 0.28 Tension escalation 2007-2012
Reputation Damage $\mu_R$ 0.50 Breach response analysis
Reputation Decay $\delta_R$ 0.02 Long-term relationship patterns
Interdependence Amp. ξ 0.45 JV structure analysis
Signal Sensitivity κ 1.0 Investment-response calibration
Initial Trust τ₀ 0.65 Strong initial relationship (2004)

Validated Value Function (TR-1)

Parameter Value Interpretation
θ 20.0 Logarithmic scale (JV value creation)
γ 0.65 Validated complementarity (panel synergies)

Reward Function

r_Samsung = π_Samsung + D_Samsung→Sony · π_Sony
r_Sony    = π_Sony    + D_Sony→Samsung · π_Samsung

Interdependence calibrated to historical mutual dependence patterns.

Episode Structure

Initial State

Historical Background

The S-LCD Joint Venture (2004-2012)

Partners:

Structure:

Outcome:

Why This Case Study?

The S-LCD venture is ideal because:

  1. Well-documented: Public financial and strategic data available
  2. Clear coopetition: Cooperation (panels) + competition (TVs)
  3. Dynamic trust: Trust evolved significantly over time
  4. Quantifiable outcomes: Investment, market share, and exit data

Environment Specification

Basic Usage

import coopetition_gym
import numpy as np

# Create environment
env = coopetition_gym.make("SLCD-v0")

obs, info = env.reset(seed=42)

# Agents are named for clarity
print(f"Agents: {env.possible_agents}")  # ['Samsung', 'Sony']

# Run episode
for step in range(100):
    # Cooperation levels for Samsung and Sony
    samsung_action = 55.0  # 55% of endowment
    sony_action = 45.0     # 45% of endowment

    actions = np.array([samsung_action, sony_action])
    obs, rewards, terminated, truncated, info = env.step(actions)

print(f"Samsung total: {sum(info.get('samsung_rewards', [0]))}")
print(f"Sony total: {sum(info.get('sony_rewards', [0]))}")

Parameters

Parameter Default Description
max_steps 100 Maximum timesteps per episode
trust_enabled True Enable trust dynamics
render_mode None Rendering mode

Agent Configuration (TR-1 §8.3)

Endowments

Agent Role Endowment Interpretation
Samsung (0) Technology Provider 100.0 Manufacturing capacity
Sony (1) Market Partner 100.0 Market access/brand

Bargaining Shares

Agent Alpha Interpretation
Samsung 0.55 Slightly higher due to technology control
Sony 0.45 Brand premium but less operational control

Samsung’s slight advantage reflects its control over manufacturing technology.

Baselines

Agent Baseline Interpretation
Samsung 30.0 30% minimum viable cooperation
Sony 30.0 30% minimum viable cooperation

Actions below baseline are considered defection.

Interdependence Structure

S-LCD Specific Dependencies

Created using create_slcd_interdependence():

D = [[ 0.00,  D_Samsung_Sony ],
     [ D_Sony_Samsung,  0.00 ]]

Validated Values:

Trust Dynamics

Parameters (TR-1 Validated)

Parameter Symbol Value Source
Trust Building Rate λ⁺ 0.08 TR-1 §8.3
Trust Erosion Rate λ⁻ 0.28 TR-1 §8.3
Reputation Damage $\mu_R$ 0.50 TR-1 §8.3
Reputation Decay $\delta_R$ 0.02 TR-1 §8.3
Interdependence Amp. ξ 0.45 TR-1 §8.3
Signal Sensitivity κ 1.0 TR-1 §8.3
Initial Trust τ₀ 0.65 Historical: Strong initial relationship

Historical Trust Evolution

The real S-LCD venture showed:

Value Function (TR-1)

Logarithmic Specification

Parameter Value Source
θ 20.0 TR-1 §8.3
γ 0.65 Validated complementarity

Value Creation

V(a_S, a_Y) = θ × ln(a_S + a_Y) × (1 + γ × complementarity)

Where a_S = Samsung’s action, a_Y = Sony’s action.

Validation Methodology

Data Collection

Sources:

60 Validation Data Points

The 60 validation data points span four categories:

Category Count Metric Tolerance
Investment levels 16 Annual JV investment (normalized) ±10%
Trust indicators 20 Contract terms, public statements, leadership interactions Qualitative match
Market outcomes 16 Market share evolution, production volumes ±5%
Exit timing 8 Relationship milestones, stake adjustments ±1 year

Matching Criteria

Investment Levels (16 points):

Trust Indicators (20 points):

Market Outcomes (16 points):

Exit Timing (8 points):

58/60 Score

Accuracy Calculation:

Match = prediction within tolerance
Score = 58/60 = 96.7% accuracy

58 matches include:

Deviations (2 points)

Two trust indicator data points showed deviation: 1. 2008 Q4 Trust Spike - External shock (global financial crisis)

  1. 2010 Q2 Trust Recovery - Unmodeled strategic event
    • Model predicted monotonic decline
    • Historical: brief recovery following executive meeting
    • Cause: Discrete strategic intervention not captured in continuous model

Sensitivity Analysis

Parameter sensitivity around validated values (±20%):

Parameter Base Range Tested Accuracy Range
λ⁺ 0.08 0.064-0.096 55-58/60
λ⁻ 0.28 0.224-0.336 54-58/60
γ 0.65 0.52-0.78 53-58/60
ξ 0.45 0.36-0.54 56-58/60

The validated parameters achieve maximum accuracy across the sensitivity range.

Confidence Intervals

Bootstrap confidence intervals (1000 samples):

Metric Point Estimate 95% CI
Overall accuracy 96.7% [93.3%, 98.3%]
Investment match 100% [93.8%, 100%]
Trust match 90% [80%, 95%]
Market match 100% [93.8%, 100%]
Exit timing 100% [87.5%, 100%]

Cross-Validation

Leave-one-out cross-validation on data categories:

Held-Out Category Accuracy on Held-Out
Investment 15/16 (93.8%)
Trust 17/20 (85.0%)
Market 15/16 (93.8%)
Exit 8/8 (100%)

The model generalizes well across categories.

Metrics and Info

The info dictionary includes:

Key Type Description
step int Current timestep
mean_trust float Average trust level
samsung_investment float Samsung’s current action
sony_investment float Sony’s current action
total_value float Total value created
samsung_payoff float Samsung’s current payoff
sony_payoff float Sony’s current payoff

Reproducing Historical Dynamics

Early Phase (High Cooperation)

# 2004-2006: Both partners invest heavily
samsung_action = 70.0  # 70% cooperation
sony_action = 65.0     # 65% cooperation

Expected outcome:

Middle Phase (Growing Tension)

# 2007-2009: Samsung invests more, Sony becomes cautious
samsung_action = 60.0  # Samsung maintains
sony_action = 45.0     # Sony pulls back

Expected outcome:

Late Phase (Pre-Exit)

# 2010-2012: Declining cooperation
samsung_action = 40.0  # Samsung reduces
sony_action = 35.0     # Sony minimizes

Expected outcome:

Example: Historical Simulation

import coopetition_gym
import numpy as np

env = coopetition_gym.make("SLCD-v0")
obs, info = env.reset(seed=42)

# Simulate historical phases
phases = [
    ("Early (2004-2006)", 30, (70.0, 65.0)),
    ("Middle (2007-2009)", 35, (60.0, 45.0)),
    ("Late (2010-2012)", 35, (40.0, 35.0))
]

total_samsung = 0
total_sony = 0

for phase_name, steps, (samsung, sony) in phases: phase_samsung = 0
    phase_sony = 0

    for step in range(steps): actions = np.array([samsung, sony])
        obs, rewards, terminated, truncated, info = env.step(actions)
        phase_samsung += rewards[0]
        phase_sony += rewards[1]

    print(f"{phase_name}:")
    print(f"  Trust: {info['mean_trust']:.3f}")
    print(f"  Samsung: {phase_samsung:.1f}, Sony: {phase_sony:.1f}")

    total_samsung += phase_samsung
    total_sony += phase_sony

print(f"\nTotal: Samsung={total_samsung:.1f}, Sony={total_sony:.1f}")

Research Applications

SLCD-v0 is the recommended environment for:

Comparison with Other Environments

Feature SLCD-v0 TrustDilemma-v0
Validation 58/60 historical Theoretical
Agents Named (Samsung, Sony) Generic (0, 1)
Parameters Fixed (validated) Default (configurable)
Trust Dynamics Moderate High
Recommended For Benchmarking Algorithm development

References

  1. Pant, V. & Yu, E. (2025). Computational Foundations for Strategic Coopetition: Formalizing Interdependence and Complementarity. arXiv:2510.18802
  2. Samsung Electronics & Sony Corporation. (2004). S-LCD Joint Venture Announcement.
  3. Ritala, P. & Hurmelinna-Laukkanen, P. (2009). What’s in it for me? Creating and appropriating value in innovation-related coopetition. Technovation.