Research

potpourri

Leela Chess Zero [www, git]

The Leela Chess Zero developers

Leela Chess Zero (Lc0) is an open source chess engine based on the AlphaZero work by DeepMind. Lc0 learns to play chess from scratch solely by playing against itself. Each iteration the engine plays games against itself across many computers (volunteers). Next we gather the data and train the next generation neural network. This network is then deployed again and the cycle repeats.

The AARTFAAC All Sky Monitor: System Design and Implementation [pdf]

API, Astron, University of Amsterdam 2016. Peeyush Prasad, Folkert Huizinga, Eric Kooistra, Daniel van der Schuur, Andre Gunst, John Romein, Mark Kuiack, Gijs Molenaar, Antonia Rowlinson, John D. Swinbank, Ralph A.M.J Wijers.

The Amsterdam-ASTRON Radio Transients Facility And Analysis Center (AARTFAAC) all sky monitor is a sensitive, real time transient detector based on the Low Frequency Array (LOFAR). It generates images of the low frequency radio sky with spatial resolution of 10s of arcmin, MHz bandwidths, and a time cadence of a few seconds, while simultaneously but independently observing with LOFAR. The image timeseries is then monitored for short and bright radio transients. On detection of a transient, a low latency trigger will be generated for LOFAR, which can interrupt its schedule to carry out follow-up observations of the trigger location at high sensitivity and resolutions. In this paper, we describe our heterogeneous, hierarchical design to manage the 240 Gbps raw data rate, and large scale computing to produce real-time images with minimum latency. We discuss the implementation of the instrumentation, its performance, and scalability.

WPSS: Watching people security services [pdf]

TNO, Cameramanager.com, VicarVision, Noldus Information Technology, Eagle Vision Systems, 2013.
Henri Bouma, Jan Baan, Sander Borsboom, Kasper van Zon, Xinghan Luo, Ben Loke, Bram Stoeller, Hans van Kuilenburg, Judith Dijk

To improve security, the number of surveillance cameras is rapidly increasing. However, the number of human operators remains limited and only a selection of the video streams are observed. Intelligent software services can help to find people quickly, evaluate their behavior and show the most relevant and deviant patterns. We present a software platform that contributes to the retrieval and observation of humans and to the analysis of their behavior. The platform consists of mono- and stereo-camera tracking, re-identification, behavioral feature computation, track analysis, behavior interpretation and visualization.

Recognising Individuals by Appearance across Non-overlapping Stereo Cameras [pdf]

University of Amsterdam, 2012.
Bram Stoeller, supervised by Leo Dorst and Gwenn Englebienne

Bram’s masters project was about color and geometry based appearance models. The project involved people recognition using non overlapping top mounted (i.e. ceiling mounted) stereo cameras. The report discusses a number of methods, including a novel radial hue-saturation histogram. It is based on the circularity of the hue-saturation space where the value is neglected. I created a method to divide a circle into equal parts. The proof that the partitions are of equal size is provided in Appendix A of the thesis.

Machine learning strategic game play for a first-person shooter video game [pdf]

Folkert Huizinga, supervised by S. Whiteson (University of Amsterdam) and R. Straatman (Guerrilla Games), 2011

The main goal of Artificial Intelligence (AI) in commercial video games is to provide a rich gaming experience for the end user. Providing this experience requires a careful balance between encoding prior knowledge, control by developers and autonomous learning AI. This thesis proposes a method that maintains a balance between these three objectives. The method proposed is a novel combination of Hierarchical Task Network (HTN) planning and Reinforcement Learning (RL).

An artificial intelligence for a realtime strategygame [code]

Folkert Huizinga, E323AI for the springrts open source engine, 2010.

Realtime strategygames are a set of strategy games where the gameplay is not turnbased like e.g. chess, but realtime. Games like Starcraft, Planetary Annihilation, Total Annihilation etc. E323AI is an A.I. that was built for the springrts engine. At the time of writing the code it was the best AI within the game with a very aggressive gameplay style until atleast 2014. To play it you can goto springrts.com, download the engine (windows or linux) and install E323AI. To observe the sourcecode click here. In 2014 two researchers K. D. Rogers and A. A. Skabar used it as a basis for an improved micromanagement system – A Micromanagement Task Allocation System for Real-Time Strategy Games.

Lisa – Layered Interactive Sokoban Assistant [pdf]

University of Amsterdam, 2010.
Bram Stoeller, Tjerk Kostelijk

This project was a literature research on solving the Japanese game Sokoban. Inspiration is emerged from previous work to develop LISA (Layered Interactive Sokoban Assistant). LISA assists the player towards solving the puzzle. AI techniques are used to display strategic information about the game is visual assistance layers.

Reconstructing a 3D Structure from Unknown Viewpoints [pdf]

University of Amsterdam, 2010.
Bram Stoeller, Tjerk Kostelijk, Folkert Huizinga

In this project we considered the problem of computing the 3D structure of an unknown arbitrarily-shaped scene from two photographs taken at unknown viewpoints using uncalibrated cameras.

AIGOA – Artificial Intelligent Go Annotator [pdf]

University of Amsterdam, 2008.
Bram Stoeller, supervised by Leo Dorst

In this project, every move in a Go game is automatically registered by a webcam. This is a completely automated process without any camera calibration. A lot of computer vision techniques are combined: background estimation, edge detection, feature detection, Hough lines etcetera.

Real-time Billboard Substitution in a Video Stream of a Soccer Match [pdf]

University of Amsterdam, 2007.
Bram Stoeller, Tjerk Kostelijk.

The goal of this project was to replace the adds at the borders of a soccer field in a live video stream with other (regional) advertisements. Further more it was the aim to do this fully automatically using computer vision, without any camera calibration.

Background Estimation

University of Amsterdam, 2006.
Bram Stoeller, Tjerk Kostelijk.

It is very convenient to have an accurate estimation of the background while tracking objects or people in a video stream. Especially in outdoor situations this brings some issues. Think about changing weather, reflections in water waving leaves in a tree etcetera. These settings ask for a solid system using a multiple gaussian based Markov decision model.

Quakevolution [pdf]

University of Amsterdam, 2006.
Bram Stoeller, Tjerk Kostelijk, Folkert Huizinga, Rik Tobi.

In our very first real AI project we created a genetic algorithm (GA) to optimize game strategies. The method: Six Quake bots play against each other and the best two are selected to breed. This iterative process repeats itself until some optimum is found. The result is indeed a bot from which no human player can ever win.

Start typing and press Enter to search