Deep learning optimization in real world applications is often limited by the lack of valuable data, either due to missing labels or the sparseness of relevant events (e.g. failures, anomalies) in the dataset. We face this problem when we optimize dispatching and rerouting decisions in the Swiss railway network, where the recorded data is variable over time and only contains a few valuable events. To overcome this deficiency we use the high computational power of modern GPUs to simulate millions of physically plausible scenarios. We use this artificial data to train our deep reinforcement learning algorithms to find and evaluate novel and optimal dispatching and rerouting strategies.
In this session we would like to demonstrate the benefits of multi-agent reinforcement learning for real world applications. We highlight our multi-agent system which is capable to learn an efficient communication protocol. The agents transmit relevant information through a low bandwidth channel to collectively solve a complex rescheduling problem. This approach proves beneficial whenever a feasible solution needs to be found within a short timeframe - in contrast to the computationally expensive optimal solution. Thus, our results can be applied to many different problems in the domain of operations research and transportation.
Key words: Deep Learning and AI, DGX, Reinforcement Learning, Traffic Management
We'll highlight the benefits of using GPU accelerated high performance simulations on DGX systems in combination with deep reinforcement learning. Deep reinforcement learning has gained a lot of momentum lately with its success in solving various computer games and control problems. Based on these promising results we pursue the approach of optimizing train schedules and train dispatching with deep reinforcement learning. We implemented GPU accelerated high performance train network simulations to allow us to train multiple realizations of the dispatching agent in parallel at super real-time speed and react to the ever-changing topology and traffic situation. The result is that we are able to perform training in a feasible time span and explore novel dispatching and scheduling strategies for current and future railway traffic. Key words: Deep Learning and AI, DGX