In the rapidly evolving landscape of reinforcement learning, selecting the right tools and platforms can fundamentally transform an organisation's ability to deliver innovative solutions. For the Technology & Innovation Team at Decision Lab, an award-winning UK technology company specialising in Mathematical Modelling, Optimisation, Simulation, Data Science and Artificial Intelligence, the adoption of AgileRL Arena marked a significant breakthrough in solving complex logistics optimisation problems through reinforcement learning.

"AgileRL Arena has significantly streamlined our RL development workflow making training and deploying agents a breeze. The platform's hyperparameter tuning capabilities have dramatically accelerated our experimentation and improved the performance of our models." — Andrew Nestor, Decision Lab T&I Team

‍

Theory-powered Practical Solutions

Decision Lab stands out in the UK technology landscape for their unique combination of academic rigour and practical implementation expertise. Their teams, composed of PhD and MSc level experts, bring together cutting-edge theoretical knowledge with hands-on experience in solving complex real-world challenges. This distinctive blend has earned them partnerships with some of the most demanding organisations in the UK, including the Defence Science and Technology Laboratory, Babcock International, and the Royal Navy.
‍
The company’s impact extends far beyond individual projects. Their solutions influence millions of people across multiple sectors, from optimising defense systems to improving healthcare delivery and revolutionising retail operations. This broad reach demands not just technical excellence, but also the ability to consistently deliver reliable, high-performance solutions that can scale across different applications and environments.

‍

Optimising Logistics Operations

Decision Lab faced a particularly challenging optimisation problem in the logistics sector: maximising the efficient use of storage space through optimal bin packing. The goal was to develop an AI system that could dynamically determine the most efficient way to pack packages of varying sizes into standardised bins, maximising space utilisation while adhering to various physical constraints and handling requirements.

This bin packing challenge represented a perfect use case for reinforcement learning, as it required the agent to learn complex spatial relationships and make sequential decisions that would optimise overall space utilisation. However, their existing solution using RLlib presented several limitations:

• ‍Hyperparameter optimisation: Finding the right balance of parameters to encourage efficient packing strategies while maintaining stable learning proved challenging with their existing tools.‍
• Resources and time: Training effective bin packing agents required significant computational resources and time, making it difficult to experiment with different approaches or scale the solution.

These limitations were particularly problematic for a company of Decision Lab's calibre, where the ability to rapidly prototype, test, and deploy sophisticated AI solutions is crucial for maintaining their competitive edge.

‍

The Implementation Journey with AgileRL

The decision to implement AgileRL Arena in October 2024 marked the beginning of a transformative period for Decision Lab's logistics optimisation capabilities. The implementation process revealed several key advantages of the platform:

• ‍Environment Integration: The team had a custom gymnasium environment that accurately modelled their bin packing challenge, including all relevant physical constraints and optimisation objectives. Using Arena, the team was able to easily upload the environment and validate whether the environment was ready for training.‍
• Support Infrastructure: The AgileRL team's customer support and responsiveness proved invaluable during the implementation phase, particularly in helping optimise the reward function for the bin packing problem and ensuring proper environment validation.‍
• Technical Architecture: The implementation process was remarkably straightforward from a technical perspective, allowing Decision Lab to focus on model training rather than infrastructure challenges.

‍

Measurable and Immediate Impact

The impact of AgileRL Arena on Decision Lab's bin packing solution was immediately apparent in several key metrics:

• ‍Utilisation performance: The platform achieved a utilisation score of approximately 0.70 - a remarkable improvement over the 0.5 scores achieved with minimal hyperparameter tuning with their previous RLlib implementation. This meant their AI system could pack 40% more into each bin.‍
• Training efficiency: The platform's evolvable hyperparameter optimisation capabilities significantly reduced the time required to achieve satisfactory model performance. The bin packing agent learned effective strategies more quickly and consistently than previous implementations. With RLlib, the agent took 1 hour 49 minutes to train. Training on Arena took just 45 minutes to achieve the same level of performance - a 60% reduction in training time.‍
• Visualisation and analysis: AgileRL Arena's comprehensive visualisation tools provided unprecedented insight into the agent's training.‍
• Deployment: Arena eliminated the need to build deployment infrastructure. The team simply had to copy the deployment code from Arena into their local script for inference.

‍

8x Concurrent model‍testing capacity

+20%‍Training performance

‍

Decision Lab's experience with AgileRL Arena demonstrates how the right platform can transform theoretical possibilities into practical advantages, enabling organisations to solve complex optimisation problems while maintaining operational efficiency and reliability. Their success in the bin packing challenge showcases the platform's ability to handle real-world logistics problems effectively, opening new possibilities for AI-driven optimisation across various industries.
‍

If you would like to take part in a case-study with AgileRL, please reach out on LinkedIn.