Institutional Repository [SANDBOX]
Technical University of Crete
EN  |  EL

Search

Browse

My Space

Deep reinforcement learning for multi-agent search and rescue operations

Chanialakis Theofilos

Full record


URI: http://purl.tuc.gr/dl/dias/72379601-0F0E-424B-A51C-6F814864B002
Year 2020
Type of Item Diploma Work
License
Details
Bibliographic Citation Theofilos Chanialakis, "Deep reinforcement learning for multi-agent search and rescue operations", Diploma Work, School of Electrical and Computer Engineering, Technical University of Crete, Chania, Greece, 2020 https://doi.org/10.26233/heallink.tuc.86822
Appears in Collections

Summary

Emergency situations, like natural disasters, can cause significant problems to our society so they require preparatory actions and immediate response to protect the population to the best of our abilities. Many groups and organizations have been established to aid in Search and Rescue and Emergency Response (ER) operations.Preparation and preparatory actions, in most cases, are not enough, so it is vital that many agencies and groups, which are specialized in ER situations, like firemen and medics, take immediate action. Collective actions and collaboration, among those groups, are essential components for Search and Rescue operations. Global knowledge of the events and the ability to evaluate the situation are major pieces in ER management. A good and quick decision can save many lives.In this thesis, we develop an administration system for Search and Rescue operations in ER situations. The system consists of two equally important and inextricable connected parts. The first part consists of the data collection and the live parameter updates. The second part pertain to decision making and task allocation to the work force in order to minimize the danger. Moreover, we provide a detailed analysis of the system's functionality and of the technologies that are responsible for the system's consistency.The system can be used by two or more administrators, simultaneously, who can markup regions which need attention. The interface is a web-page with the use of augmented map and additional graphics to help with the system handling. The positions of the work-forces groups have been added and are updated frequently to the spatial data of the map. These live updates are possible due to an app which we developed for smartphones.Decision making procedure makes use of the above information and allocate tasks to every group. Machine Learning algorithms in Multi-Agent Systems/Environments are added in the system in order to make better decisions. In particular, Reinforcement Learning and Deep Neural Network architectures are combined to make sure that the actions are near optimal and the task allocation is the most efficient. Deep Reinforcement Learning is a state-of-the-art technique and it is very interesting to explore how it could be used in Multi-Agent environments with high complexity. In this thesis, we propose a novel Deep Reinforcement Learning architecture in Multi-Agent Settings, giving solutions to many problems which Machine Learning has difficulty to handle. We also provide experimental results, which indicate that the system gradually learns in realistic situations, generating meaningful action plans for all the agents.

Available Files

Services

Statistics