Deep Reinforcement Learning Applied to the Game of Qwixx

Share

A technical program starting at 6:30PM  with refreshments and networking from 6:00 to 6:30.

Technical Abstract:  Reinforcement learning methods have been steadily gaining popularity within the machine-learning community as an approach to learn gaming strategy through trial and error. One of the downsides of classical reinforcement learning is the limited applicability to games with a large state spaces which has resulted in the adoption of deep learning methods to approximate value functions defined over the state. However learning can still be difficult for complex games, particularly for those with a large stochastic component that can results in similar strategies having very different outcomes. In this work a deep reinforcement learning algorithm is presented to address these issues as seen in a representative dice based multi-player game known as Qwixx. Without knowing any rules about the game the algorithm eventually learns winning strategies by playing against a human generated method based on optimal control. 



  Date and Time

  Location

  Contact

  Registration



  • Advanced Test Equipment Rentals
  • 10401 Roselle Street
  • San Diego, California
  • United States 92121
  • Click here for Map
  • Stephen Stubberud - scstubberud@ieee.org

  • Co-sponsored by CH06204 - San Diego Section Chapter, CIS11
  • Starts 23 April 2019 08:00 AM
  • Ends 14 May 2019 10:00 PM
  • All times are America/Los_Angeles
  • No Admission Charge
  • Register


  Speakers

Dr. Gideon Prior of General Atomics

Topic:

Deep Reinforcement Learn Appled to the Game of Qwixx

Reinforcement learning methods have been steadily gaining popularity within the machine-learning community as an approach to learn gaming strategy through trial and error. One of the downsides of classical reinforcement learning is the limited applicability to games with a large state spaces which has resulted in the adoption of deep learning methods to approximate value functions defined over the state. However learning can still be difficult for complex games, particularly for those with a large stochastic component that can results in similar strategies having very different outcomes. In this work a deep reinforcement learning algorithm is presented to address these issues as seen in a representative dice based multi-player game known as Qwixx. Without knowing any rules about the game the algorithm eventually learns winning strategies by playing against a human generated method based on optimal control. 

Biography:

Gideon Prior was born in Los Angeles California in 1975. He received a B.S. (2007), M.S (2009) and Ph.D. (2013) degrees in electrical engineering from the University of California, San Diego. He joined General Atomics in La Jolla, California in 2006 where he has worked in research and development on projects including axial flux motor control, autonomous underwater navigation, aircraft arresting gears, high voltage DC breakers, hypersonic missile trajectory estimation and prediction and long distance high energy laser control.

His research interests include reinforcement learning, evolutionary algorithms, recurrent neural networks, optimal control, switched input systems and power electronics.