📞 Optimising Call Centre Operations with Reinforcement Learning (PPO)

University of Adelaide – Master’s Final Year Project
Supervisor: Dr. Wathsala Karunarathne

---

🚀 Project Overview

Traditional call centre routing systems rely on static, rule-based allocation, often leading to:

• ⏳ Longer wait times
• ❌ Lower first-call resolution rates
• 😡 Declining customer satisfaction

This project builds a Reinforcement Learning (RL)-based solution using Proximal Policy Optimisation (PPO) to dynamically assign calls to agents. Unlike static rules, the RL agent continuously adapts to:

• Real-time queue conditions
• Agent expertise and availability
• Call complexity and topic variation

---

🔑 Key Results

✅ 18% improvement in customer satisfaction
✅ 15% increase in first-call resolution rates
✅ Average wait time reduced to 54.5s (vs. 67s baseline)
✅ Maintained strong performance during peak call loads
✅ Achieved improvements without extra staffing costs

⚡ This proves reinforcement learning can optimise large-scale service systems in the real world.

---

🛠️ Technical Stack

• Languages: Python (Pandas, NumPy, Scikit-learn)
• RL Frameworks: PyTorch, Stable Baselines3
• Simulation: Custom OpenAI Gym environment (CallCenterEnv)
• Visualisation: Matplotlib, Seaborn
• Data: 10,000+ historical call centre records (ModifiedCallCenter.xlsx)

---

🧠 Methodology

1. Data Engineering

   • Cleaned 10k+ call logs
   • Engineered features: Call Congestion Index, Hourly Call Trends, Agent Performance Tiers

2. Environment Design

   • Built a custom Gym-compatible environment simulating queue lengths, call topics, and agent efficiency

3. Reward Function

   • Balanced wait time, resolution success, satisfaction scores, agent-topic match

4. PPO Agent Training

   • Actor-Critic networks (2×256 units, ReLU)
   • Trained for 200+ episodes
   • Tuned hyperparameters (γ=0.99, λ=0.95, ε=0.2, lr=3e-4)

5. Evaluation

   • Benchmarked against rule-based routing
   • Tested across peak hours & topic complexities

---

📊 Performance Summary

Metric	Baseline	PPO Agent	Δ Improvement
Avg Wait Time (sec)	67	54.5	↓ 19%
First Call Resolution (FCR)	0.63	0.73	+15%
Customer Satisfaction (1–5)	2.5	2.95	+18%
90th Percentile Wait (sec)	140	111	↓ 21%

---

📈 Key Visuals

Customer Wait Time Reduction

Agent Satisfaction by Topic

🌍 Business Impact

• Customer Experience → Faster response + higher satisfaction → stronger retention.
• Operational Efficiency → Better agent allocation without new hires.
• Scalability → PPO adapts to real-time spikes (weekends, seasonal campaigns).
• Transferability → Approach generalises to banking hotlines, healthcare helplines, and IT service desks.

---

📂 Repository Structure

├── CallCenterPPO.ipynb # End-to-end training + evaluation
├── environment.py # Custom OpenAI Gym call centre environment
├── results/ # Visualisations + evaluation plots
└── README.md # Project overview

---

▶️ How to Run

# Clone repo
git clone https://github.com/coderharry1/CallCenterPPO.git
cd CallCenterPPO

# Install dependencies
pip install -r requirements.txt

# Run PPO training
python CallCenterPPO.ipynb

---

📌 Future Directions

• Add customer segmentation for personalised routing
• Deploy prototype in cloud-native environments (AWS / GCP)
• Extend to multi-agent reinforcement learning for distributed call centres

---

🙏 Acknowledgements

Special thanks to Dr. Wathsala Karunarathne for her guidance in reinforcement learning and optimisation, and to the University of Adelaide’s School of Mathematical Sciences for academic support.

🔗 Links

👉 Explore the full code + training logs here: CallCenterPPO.ipynb

🎯 Role Relevance Statement

This project proves I can combine SQL, Python, and Machine Learning to deliver measurable business impact — the same skills I’ll bring to a Data Analyst / BI Analyst role.