A sophisticated multi-objective optimization system for hotel ranking and pricing strategies, built with FastAPI backend and R Shiny frontend.

• Multi-Objective Optimization: Balance Trivago revenue, user satisfaction, and partner value
• Reinforcement Learning: Pre-trained DQN agent for optimal policy selection
• Bandit Simulation: Comprehensive click-through rate analysis
• Strategic Simulation: End-to-end testing of optimization strategies
• Data Generation: Realistic hotel and user data generation
• Ecosystem Health Monitoring: Budget utilization and market balance analysis

• Backend: FastAPI (Python) with PuLP optimization
• Frontend: R Shiny with interactive visualizations
• Containerization: Docker Compose for easy deployment
• Data Storage: CSV files for persistence and analysis

graph TB
    subgraph "Frontend Layer"
        UI[R Shiny UI]
        Viz[Interactive Visualizations]
    end
    
    subgraph "Backend Layer"
        API[FastAPI Server]
        Opt[PuLP Optimizer]
        RL[DQN Agent]
        Data[Data Generator]
    end
    
    subgraph "Data Layer"
        CSV[CSV Files]
        Model[RL Model]
    end
    
    subgraph "External Systems"
        Market[Market Data]
        Users[User Behavior]
        Partners[Partner Data]
    end
    
    UI --> API
    API --> Opt
    API --> RL
    API --> Data
    Opt --> CSV
    RL --> Model
    Data --> CSV
    Market --> Data
    Users --> Data
    Partners --> Data
    CSV --> Viz
    Model --> RL

The system uses a pre-trained Deep Q-Network (DQN) agent with four distinct policies:

The DQN observes a 6-dimensional state vector representing current market conditions. Each dimension is normalized to the [0,1] range for optimal neural network performance.

Market Demand (0-1)

• High (0.8-1.0): Strong market activity, many users searching
• Medium (0.4-0.7): Moderate market activity
• Low (0.0-0.3): Weak market activity, few users

Days to Go (0-1)

• Low (0.0-0.2): Immediate bookings (0-73 days)
• Medium (0.2-0.5): Short-term bookings (73-183 days)
• High (0.5-1.0): Long-term bookings (183+ days)

Competition Density (0-1)

• Low (0.0-0.3): Monopoly/oligopoly (1-6 partners)
• Medium (0.3-0.7): Competitive market (6-14 partners)
• High (0.7-1.0): Highly competitive (14+ partners)

Price Volatility (0-1)

• Low (0.0-0.3): Stable pricing, predictable market
• Medium (0.3-0.7): Moderate price fluctuations
• High (0.7-1.0): High price volatility, dynamic market

Satisfaction Trend (0-1)

• Low (0.0-0.3): Declining user satisfaction
• Medium (0.3-0.7): Stable satisfaction
• High (0.7-1.0): Improving user satisfaction

Budget Utilization (0-1)

• Low (0.0-0.3): Under-utilized budgets (0-30%)
• Medium (0.3-0.7): Balanced budget usage (30-70%)
• High (0.7-1.0): High budget utilization (70-100%)

The DQN can choose from 4 predefined policies:

The reward is calculated based on optimization results using a multi-objective approach that balances all three business objectives.

$$R = 0.4 \times \frac{\text{Trivago Income}}{1000} + 0.3 \times \frac{\text{User Satisfaction}}{10} + 0.3 \times \frac{\text{Partner Conversion Value}}{10}$$

Where:

• Trivago Income: Revenue generated for Trivago (normalized by 1000)
• User Satisfaction: User satisfaction score (normalized by 10)
• Partner Conversion Value: Value generated for partners (normalized by 10)

The DQN uses an epsilon-greedy strategy to balance exploration and exploitation:

Exploration Phase (ε = 1.0 → 0.01):

• Random policy selection to discover new strategies
• Gradually reduces exploration rate as learning progresses

Exploitation Phase (ε ≈ 0.01):

• Uses learned Q-values to select optimal policies
• Maximizes expected reward based on experience

The DQN uses a feedforward neural network with the following structure:

The Q-learning algorithm updates Q-values using the Bellman equation:

$$Q(s_t, a_t) \leftarrow Q(s_t, a_t) + \alpha \left[ r_t + \gamma \max_{a'} Q(s_{t+1}, a') - Q(s_t, a_t) \right]$$

Where:

• $$s_t$$: Current state
• $$a_t$$: Selected action
• $$r_t$$: Received reward
• $$s_{t+1}$$: Next state
• $$\alpha$$: Learning rate
• $$\gamma$$: Discount factor (0.95)

The DQN learns market-condition → optimal-policy mappings through experience. The agent discovers patterns that maximize long-term rewards across different market scenarios.

Example 1: High-Demand, Low-Competition Market

• State Vector: [0.8, 0.1, 0.2, 0.3, 0.6, 0.4]
• Market Interpretation:
  • High demand (80 offers)
  • Immediate bookings (37 days average)
  • Low competition (4 partners)
  • Low price volatility
  • Improving satisfaction
  • Medium budget utilization
• Expected Policy: High-Revenue Policy

Example 2: Low-Demand, High-Competition Market

• State Vector: [0.3, 0.8, 0.9, 0.7, 0.4, 0.8]
• Market Interpretation:
  • Low demand (30 offers)
  • Long-term bookings (292 days average)
  • High competition (18 partners)
  • High price volatility
  • Declining satisfaction
  • High budget utilization
• Expected Policy: High-Trust Policy

Target Network Updates: Every 100 steps

• Prevents overestimation of Q-values
• Stabilizes training convergence
• Improves learning stability

Experience Replay: Batch size of 32

• Breaks temporal correlations
• Improves sample efficiency
• Enables stable learning

1. Adaptive Learning: DQN adapts to changing market conditions
2. Multi-Objective Optimization: Balances revenue, user satisfaction, and partner value
3. Experience Replay: Learns from past experiences to improve future decisions
4. Exploration vs Exploitation: Balances trying new strategies vs using proven ones
5. Market-Aware Policies: Selects policies based on current market state

The DQN essentially learns "Which policy works best in which market condition?" through trial and error, optimizing for long-term performance across all business objectives!

flowchart TD
    A[Market State Observation] --> B[State Vector Computation]
    B --> C[6-Dimensional State Vector]
    C --> D[Neural Network Processing]
    D --> E[Q-Value Calculation]
    E --> F[Policy Selection]
    F --> G[Action Execution]
    G --> H[Optimization Execution]
    H --> I[Reward Calculation]
    I --> J[Experience Storage]
    J --> K[Network Training]
    K --> L[Q-Value Update]
    L --> M[Next State Observation]
    M --> A
    
    subgraph "State Vector Components"
        S1[Market Demand]
        S2[Days to Go]
        S3[Competition Density]
        S4[Price Volatility]
        S5[Satisfaction Trend]
        S6[Budget Utilization]
    end
    
    subgraph "Policy Options"
        P1[High-Trust Policy]
        P2[Balanced Policy]
        P3[High-Revenue Policy]
        P4[Partner-Focused Policy]
    end
    
    C --> S1
    C --> S2
    C --> S3
    C --> S4
    C --> S5
    C --> S6
    F --> P1
    F --> P2
    F --> P3
    F --> P4

The RL agent analyzes market conditions to select optimal policies:

Market State Variables:

• Market Demand: Number of available offers (normalized 0-1)
• Days to Go: Average booking lead time (normalized 0-1)
• Competition Density: Number of unique partners (normalized 0-1)
• Price Volatility: Standard deviation of prices (normalized 0-1)
• Satisfaction Trend: User satisfaction trend (-1 to 1, normalized to 0-1)
• Budget Utilization: Partner budget consumption percentage (0-1)

Policy Selection Logic:

• High-Trust Policy: Selected when user satisfaction is critical (high competition, low user satisfaction)
• Balanced Policy: Default choice for normal market conditions
• High-Revenue Policy: Chosen when revenue opportunities are high (high demand, low competition)
• Partner-Focused Policy: Applied when partner relationships are priority (high budget utilization, low partner satisfaction)

Stage 1: Optimal Ranking for Click Maximization

Primary Objective Function:

$$\text{Maximize: } \alpha \times \text{Trivago_Income} + \beta \times \text{User_Satisfaction} + \gamma \times \text{Partner_Conversion_Value}$$

Where:

• Trivago_Income: $$\sum_{i,j} (\text{CTR}i \times pConvert_j \times Commission_j \times Price_j \times X{ij})$$
• User_Satisfaction: $$\frac{\sum_{i,j} (\text{CTR}i \times Satisfaction_j \times X{ij})}{\sum_{i,j} (\text{CTR}i \times X{ij})}$$ (weighted average, 0-10 scale)
• Partner_Conversion_Value: $$\sum_{i,j} (\text{CTR}i \times pConvert_j \times Price_j \times X{ij})$$

Stage 1 Constraints:

Assignment Constraints:

• $$\sum_j X_{ij} \leq 1$$ for each position $$i$$ (each position at most one offer)
• $$\sum_i X_{ij} \leq 1$$ for each offer $$j$$ (each offer at most one position)

Budget Constraints:

• $$\sum_{i,j} (\text{CTR}_i \times \text{CPC}j \times X{ij}) \leq \text{Remaining_Budget}_P$$ for each partner $$P$$

Weight Constraints:

• $$\alpha + \beta + \gamma = 1$$ (weights sum to unity)
• $$\alpha, \beta, \gamma \geq 0$$ (non-negative weights)

Stage 2: Offer Hiding for Reconversion & Budget Rationalization

Hiding Decision Function:

• Hide offer $$j$$ if: $$\text{Reconversion_Probability}_j < \text{Threshold}$$ (default: 0.3)
• Hide offer $$j$$ if: $$\text{Budget_Utilization} > \text{Target}$$ (default: 0.8)

Budget Utilization Constraint:

• $$\frac{\sum_{j \in \text{Visible}} (\text{Expected_Clicks}_j \times \text{CPC}_j)}{\text{Total_Budget}_P} \leq \text{Target_Utilization}$$

Stage 1 Variables:

• $$X_{ij}$$: Binary variable indicating if offer $$j$$ is placed at position $$i$$ for user $$u$$
• $$\alpha, \beta, \gamma$$: Weight parameters for multi-objective optimization

Stage 2 Variables:

• $$H_j$$: Binary variable indicating if offer $$j$$ is hidden (1 = hidden, 0 = visible)

$$\text{CTR}(\text{position}) = \frac{1}{1 + 0.3 \times \text{position}}$$

$$\text{Reconversion_Probability}_j = 0.7 \times \text{Conversion_Probability}_j$$

Trivago Income: Total revenue from commissions and conversions User Satisfaction: Weighted average satisfaction score (0-10 scale) Partner Conversion Value: Total value generated for partners Total Objective: $$\alpha \times \text{Trivago_Income} + \beta \times \text{User_Satisfaction} + \gamma \times \text{Partner_Conversion_Value}$$

1. Data Generation: Create realistic hotel and user datasets
2. Strategic Simulation: Test optimization with current weights
3. Policy Selection: RL agent chooses optimal policy based on market conditions
4. Optimization: Multi-objective ranking optimization
5. Results Analysis: CSV export for ecosystem health and causal impact analysis

# Clone the repository
git clone <repository-url>
cd Dynamic-Pricing-app

# Start the application
docker-compose up -d

• Frontend: http://localhost:3838
• Backend API: http://localhost:8001
• API Documentation: http://localhost:8001/docs

Key data files in /data:

• trial_sampled_offers.csv: Main dataset for optimization
• bandit_simulation_results.csv: Click-through rate analysis
• optimization_ranking_results.csv: Optimal hotel rankings
• optimization_objectives_results.csv: Objective function values
• dqn_model.pth: Pre-trained RL model

Adjust optimization parameters in the Strategic Levers tab:

• α (Alpha): Trivago revenue weight
• β (Beta): User satisfaction weight
• γ (Gamma): Partner value weight

Monitor system performance through:

• Ecosystem Health: Budget utilization, market balance
• Causal Impact: Shapley values, counterfactual analysis
• Strategic Simulation: End-to-end optimization results
• Bandit Analysis: Click-through rate optimization

1. Fork the repository
2. Create a feature branch
3. Make your changes
4. Submit a pull request

This project is licensed under the MIT License.

The linear programming model implementation in backend/main.py is fully coherent with the mathematical formulations documented in this README:

✅ Objective Function Components:

• Trivago Income: CTR_i × pConvert_j × Commission_j × Price_j ✓
• User Satisfaction: CTR_i × Satisfaction_j (normalized to weighted average) ✓
• Partner Conversion Value: CTR_i × pConvert_j × Price_j ✓

✅ Constraints Implementation:

• Assignment Constraints: ∑_j X_ij ≤ 1 and ∑_i X_ij ≤ 1 ✓
• Budget Constraints: ∑_{i,j} (CTR_i × CPC_j × X_ij) ≤ Remaining_Budget_P ✓
• Weight Constraints: α + β + γ = 1 and α, β, γ ≥ 0 ✓

✅ Hiding Decision Logic:

• Reconversion Threshold: Reconversion_Probability_j < 0.3 ✓
• Budget Utilization: Budget_Utilization > 0.8 ✓
• Reconversion Formula: 0.7 × Conversion_Probability_j ✓

✅ CTR Formula:

• CTR(position) = 1 / (1 + 0.3 × position) ✓

✅ Weighted Average User Satisfaction:

• User_Satisfaction = Σ(CTR_i × Satisfaction_j × X_ij) / Σ(CTR_i × X_ij) ✓

✅ Total Objective:

• α × Trivago_Income + β × User_Satisfaction + γ × Partner_Conversion_Value ✓

✅ CSV Export Structure:

• All objective function values calculated per README specifications ✓
• Individual row calculations match mathematical formulations ✓
• Weighted averages properly implemented ✓

The implementation ensures mathematical rigor while maintaining computational efficiency through PuLP optimization.