Reinforcement learning (RL) has achieved remarkable outcomes in real-world settings in large tech companies including Alphabet, Amazon, Meta and Microsoft. The challenge is to design the software systems to train RL models and heterogeneous data at a very large scale. For instance, training GPT-4, the large language model behind the popular chatbot ChatGPT, requires the distribution of the model and data across tens of thousands of special hardware, graphics processing units (GPU). As a result, it becomes difficult or even impossible for common users and small and medium enterprises to apply modern RL frameworks in their actual business. In this project carried out at Aalto University, we will design and build a scalable and flexible RL framework, AthenaRL, at the industrial scale. AthenaRL will be open-sourced with ease-to-use interfaces and an end-to-end deployment pipeline. Therefore, users can directly use or customize AthenaRL to solve their own domain-specific problems.

2024

2028