Helyszín: Online, https://meet.google.com/pfp-yzpu-cbe?authuser=0
Időpont: 2020.06.30., 17:30 – 17:50
Kivonat: Applying reinforcement learning (RL) methods is a difficult task since these methods usually require a standardization of their inputs and in a similar way a standardization of the output – action – spaces. Usually, when applied to real-world problems, the success of the methods is highly dependent on the particularities of the representation and the synergies between the input representation and the problem we are solving – manifested e.g. trough the decision function. The application of the reinforcement learning for rail optimization is difficult since the input space is huge: next to the position of the trains on the graph of the rail tracks, the topology itself – on which the trains operate – is required. In this project we aim to set up a modularized solution to the scheduling problem within a fairly crowded rail system. We will be using the simulation environment provided by the FLATLAND challenge (www.aicrowd.com flatland-challenge) within the Python programming system. Investigations into the project have already been started – they ended in January 2020 – and the conclusions were that a fair mount of the dynamic re-scheduling can be done using classical and slightly improved graph search algorithms and deadlock avoidance strategies. When “intelligence was intended”, special attention was required when coding input: all of the tested out-of-the-box methods failed for this task. The aim of the project is to look for a better representation for the scheduling problem and to try to shape the input representation in a way that is both (1) explainable to humans and (2) successful for the RL algorithm.