Check out Giacomo's latest work
New Open Access publication now out in Machine Learning (Springer Nature Group): POMDP inference and robust solution via deep reinforcement learning: an application to railway optimal maintenance.
In this work, led by Giacomo Arcieri (Chair of Structural Mechanics and Monitoring at ETH Zurich), we propose a joint framework of POMDP inference and robust solution based on MCMC sampling, deep RL, and domain randomization. The proposed framework is applied to the real-world problem of railway optimal maintenance based on monitoring data. Two state-of-the-art model-free RL methods, which are based on LSTMs and Transformers and work directly on the observation space, are compared with a belief-input RL method. The latter is able to outperform the alternatives and significantly improves over the real-life policy.
The work forms a collaboration with Cyprien Hoelzl (irmos technologies AG and former SMM member), Eleni Chatzi from the Chair of Structural Mechanics and Monitoring at ETH Zurich, Oliver Schwery (SBB CFF FFS), Daniel Straub (Technical University of Munich), and Konstantinos Papakonstantinou (Penn State University).
The work is supported by the Swiss Federal Railways (SBB) as part of the ETH Mobility Initiative project REASSESS, which is coordinated by the Centre for Sustainable Future Mobility - ETH Zurich. We further thank David Haener and Evelyn Weiss (SBB CFF FFS) for their immense contributions to the REASSESS project.
Read more in external page Machine Learning