We explore how Learned Models and Monte Carlo Tree Search can be used in the Combinatorial Optimization problem of train rescheduling. More exactly, we propose adapting DeepMind's MuZero algorithm to offer a scalable and time-efficient solution for train rescheduling from a single-agent modeling perspective. Our method leverages on a permutation invariant representation function to build, based on the trains trajectory, the root node of the MCTS. This resusts in a learned representation that captures the complex dynamics of a railway system and able to explore its large action space.

Stroke lesion segmentation is crucial for diagnosing and treating stroke patients. Manual segmentation is labor-intensive and time-consuming for radiologists. Current deep learning methods for automation have limitations, as they don't fully exploit 3D spatial coherence or MRI modalities' complementary nature. We present a novel method, integrating T1-weighted and FLAIR MRI modalities through a custom algorithm for stroke lesion segmentation. Our approach pretrains the model on a large dataset, then fine-tunes it using in-house data with T1 and FLAIR modalities.

In most IoT networks, end devices must perform a centralized identification in order to connect to the network. This identification step consists of verifying that the end device is authorized to connect using its identifier. To decentralize this step, instead of using a standard blockchain (Bitcoin or Ethereum), we propose a new approach using a proof of identification (PoI) to measure the degree of trust of the nodes. Below a predefined threshold, a node is banned from the blockchain network while keeping its status for the IoT network.