Some Talks by Members of my Research Group
some of our talks on algorithms in combinatorial games.
Memory-Augmented Monte Carlo Tree Search
Talk given by
C. Xiao, Feb 6,
AAAI 2018 New Orleans, outstanding paper award.
Created: July 10, 2011 Last modified: May 11, 2016
The Two-edged Nature of Diverse Action Costs
Talk given by Gaojian Fan,
June 21, ICAPS 2017 Pittsburgh.
Heuristic Search in the Age of Deep Learning
Talk given by Martin Müller, May 28, GMIS Beijing.
Computer Go: from the Beginnings to AlphaGo
Revised version of talks given by Martin Müller, June 2 NUDT, Changsha, June 4, Tsinghua University, and June 5, Baidu Research.
Abstract: The ancient Chinese game of Go, or Wei Qi, has been used as a test bed for measuring progress in Artificial Intelligence for decades. After long periods of slow improvement, two relatively recent advances in heuristic search have led to rapid progress, culminating in Deepmind's development of their famous AlphaGo program. This lecture will explain the most important components in the creation of AlphaGo, and discuss its recent match against World #1 player Ke Jie from a Computing Science perspective.
Modern Heuristic Search: Towards a Unifying Framework
Revised version of talks given by Martin Müller, May 31, Tencent AI Lab, Shenzhen, and June 3, NUDT, Changsha.
AlphaGo is the most prominent example of many successful applications of modern heuristic search. The common novel architectural theme of these programs is that they combine all the following four elements: deep search, knowledge, simulations, and machine learning. In contrast, previous successes, such as IBM's chess machine Deep Blue, which famously defeated Garry Kasparov exactly 20 years ago, focused on just two of those components: search, and human-designed domain-specific knowledge. This lecture outlines a framework for modern heuristic search, showing many examples from diverse application areas which combine the four elements above. While we now have many successful case studies, much fundamental research remains to be done to understand how such systems work, and how to design new ones. In the last part of this talk, I will summarize some lessons learned, and outline preliminary ideas that could lead to a general framework for modern heuristic search.
Mastering the Game of Go: Can a Computer Program Beat a Human Champion?
"General audience" talk given by Martin Müller, March 14, 2016 at Faculty of Science, CCIS 1-140.
Note that this talk was given a few hours before game 5 of the Lee Sedol-AlphaGo match.
Computer Go Research - The Challenges Ahead.
Keynote given by Martin Müller at
IEEE CIG 2015,
Tainan, Taiwan, September 2015.
Using Domain-specific Knowledge for Monte Carlo Tree Search in Go.
Seminar talk given by Martin Müller at National Chiao Tung University,
Hsinchu, Taiwan, August 2015.
Continuous Arvand: Motion Planning with Monte Carlo Random Walks
Weifeng Chen and Martin Müller.
Workshop on Planning and Robotics.
Talk co-developed and presented by Robert Holte.
From Deep Blue to Monte Carlo
Full-day tutorial on game tree search,
Akihiro Kishimoto and Martin Müller, given at AAAI-14 in Quebec City, July 2014.
Random Walk Planning: Theory, Practice, and Application
CAIAC best PhD thesis award talk,
invited talk given by Hootan Nakhost.
Canadian Conference on Artificial Intelligence, Waterloo,
Monte Carlo Tree Search and Computer Go
University of Alberta, Cmput 366 guest lecture, given by Martin Müller, December 2011.
Planning with Monte Carlo Random Walks: New Results (8 MB pdf file)
Talk by Martin Müller, joint work with Hootan Nakhost and Fan Xie.
Seminar, Université Paris Dauphine, June 21 (1 hr), 2011, and
Université Paris Sud, June 22, 2011 (1 hr).
Abstract: Tree search using Monte Carlo simulations has been very successful in games such as Go.
such as UCT have highlighted the importance of the tradeoff between exploration and
exploitation in tree search. In our work on classical,
domain-independent planning, we have investigated randomized exploration
techniques which are inspired by Monte Carlo Tree Search.
In contrast to most other state of the art approaches, which focus
on exploiting states with low heuristic evaluation, our Arvand series of
planners includes a large exploration component through sequences of random actions.
I will explain the basic idea of planning using random walks,
and discuss several recent improvements and applications, such as combining
random walks with local tree search and "smart restarts".
I will also include a brief report of the results of the planning competition,
which will have been announced at the ICAPS conference in mid June.
Challenges in Monte Carlo Tree Search
Symposium: Driven by Search. Universiteit Maastricht, Department of Knowledge Engineering (DKE),
Talk by Martin Müller, May 24, 2011 (45 min)
Monte-Carlo tree search has revolutionized computer Go and is having an ever
increasing impact on other games and applications. However, in its current form
it does not solve all problems in difficult domains such as Go.
In 2010, a man-machine competition was held in Barcelona, Spain.
We use an analysis of the games played by our program Fuego-GB Prototype
in this event to highlight some of the challenges that need to be addressed in the future,
such as scaling to larger problems, improved simulations and integrating local analysis
The game of Go, Monte Carlo Tree Search and Computer Go
Guest lecture by Martin Müller, March 9, 2011, University of Alberta, Dept. of Philosophy, in Prof.
John Simpson's Game Theory course (90 min.)
An introductory talk about the topic of computer Go and the very successful
approaches based on simulated games and Monte Carlo Tree Search.