Gal A. Kaminka: Publications

Sorted by DateClassified by Publication TypeClassified by TopicGrouped by Student (current)Grouped by Former Students

Reinforcement Learning in Multi-Robot Swarms

Yinon Douchan. Reinforcement Learning in Multi-Robot Swarms. Master's Thesis, Tel Aviv University,2018.

Download

[PDF]4.5MB  

Abstract

In multi-robot systems, robots cannot act without interactions and conflicts must be resolved. Such conflict is a spatial conflict; robots cannot share the same spot at the same time and must avoid collision. While there are many approaches to collision avoidance and resolution, every approach has its advantages and disadvantages. Recent promising work tries to address multi robot spatial coordination by adaptive selection of reactive coordination methods using an intrinsic reward function named the Effectiveness Index (EI) which is based on the resource spending rate of the robots. While it has many desirable characteristics in terms of multi-robot coordination and shows some empirical success, its success is only limited and there are no theoretical guarantees of optimality or convergence, as we indeed show. The contributions of this work are in several areas: First, we start by listing gaps between existing theory and practice that rise in the context of reactive arbitration of coordination methods. Then, we give theoretical modelling and practical solutions in order to bridge those gaps. The theoretical modelling starts by representing a task run as an extensive form game. It then goes to creating a connection between the system-wide performance for a task run and the choices of each robot in each collision. Finally, it deals with the issue of how robots should act in order to achieve optimal system-wide performance. The practical solutions further bridge gaps that rise from running multi-robot systems in real-world applications. The last part of this work puts the theoretical modelling and practical solutions to the test by experimenting with different multi-robot domains both with real robots and in a simulation.

Additional Information

BibTeX

@mastersthesis{douchan-msc,
  author = {Yinon Douchan},
  title = {Reinforcement Learning in Multi-Robot Swarms},
  school = {{T}el {A}viv {U}niversity},
  year = {2018},
  OPTkey = {},
  OPTtype = {},
  OPTaddress = {},
  OPTmonth = {},
  OPTnote = {Available at \url{http://www.cs.biu.ac.il/~galk/Publications/b2hd-douchan-msc.html}.},
  OPTannote = {},
  wwwnote = {}, 
  abstract = {In multi-robot systems, robots cannot act without interactions and conflicts must be resolved. Such  conflict is a spatial conflict; robots cannot share the same spot at the same time and must avoid collision. While  there are many approaches to collision avoidance and resolution, every approach has its advantages and  disadvantages. Recent promising work tries to address multi robot spatial coordination by adaptive selection of  reactive coordination methods using an intrinsic reward function named the \emph{Effectiveness Index (EI)} which  is based on the resource spending rate of the robots. While it has many desirable characteristics in terms of  multi-robot coordination and shows some empirical success, its success is only limited and there are no  theoretical guarantees of optimality or convergence, as we indeed show. The contributions of this work are in  several areas: First, we start by listing gaps between existing theory and practice that rise in the context of  reactive arbitration of coordination methods. Then, we give theoretical modelling and practical solutions in order  to bridge those gaps. The theoretical modelling starts by representing a task run as an extensive form  game. It then goes to creating a connection between the system-wide performance for a task run and the choices of   each robot in each collision. Finally, it deals with the issue of how robots should act in order to achieve optimal system-wide performance. The practical solutions further bridge gaps that rise from running multi-robot   systems in real-world applications. The last part of this work puts the theoretical modelling and practical  solutions to the test by experimenting with different multi-robot domains both with real robots and in a  simulation.
  },
}

Generated by bib2html.pl (written by Patrick Riley ) on Fri Aug 30, 2024 17:29:51