Dapeng Zhang – Publications

(Show all abstracts) (Hide all abstracts)

2011

Cai Zhongjie, Dapeng Zhang and Bernhard Nebel.
Playing Tetris Using Bandit-Based Monte-Carlo Planning.
In Proceedings of AISB 2011 Symposium: AI and Games (AISB 2011). 2011.
(Show abstract) (Hide abstract) (PDF)

Tetris is a stochastic, open-ended board game. Existing artificial Tetris players often use different evaluation functions and plan for only one or two pieces in advance. In this paper, we developed an artificial player for Tetris using the bandit-based Monte-Carlo planning method (UCT). In Tetris, game states are often revisited. However, UCT does not keep the information of the game states explored in the previous planning episodes. We created a method to store such information for our player in a specially designed database to guide its future planning process. The planner for Tetris has a high branching factor. To improve game performance, we created a method to prune the planning tree and lower the branching factor. The experiment results show that our player can successfully play Tetris, and the performance of our player is improved as the number of the played games increases. The player can defeat a benchmark player with high probabilities.
Dapeng Zhang and Bernhard Nebel.
Feature Induction of Linear-Chain Conditional Random Fields - A Study Based on a Simulation.
In Proceedings of the 3rd International Conference on Agents and Artificial Intelligence (ICAART 2011). 2011.
(Show abstract) (Hide abstract) (PDF)

Conditional Random Fields (CRFs) is a probabilistic framework for labeling sequential data. Several approaches were developed to automatically induce features for CRFs. They have been successfully applied in real-world applications, e.g. in natural language processing. The work described in this paper was originally motivated by processing the sequence data of table soccer games. As labeling such data is very time consuming, we developed a sequence generator (simulation), which creates an extra phase to explore several basic issues of the feature induction of linear-chain CRFs. First, we generated data sets with different configurations of overlapped and conjunct atomic features, and discussed how these factors affect the induction. Then, a reduction step was integrated into the induction which maintained the prediction accuracy and saved the computational power. Finally, we developed an approach which consists of a queue of CRFs. The experiments show that the CRF queue achieves better results on the data sets in all the configurations.

2010

Dapeng Zhang, Cai Zhongjie and Bernhard Nebel.
Playing Tetris Using Learning by Imitation.
In Proceedings of the 11th annual European Conference on Simulation and AI in Computer Games (GAMEON 2010). 2010.
(Show abstract) (Hide abstract) (PDF)

Tetris is a stochastic and open-end board game. Several artificial players were developed to automatically play Tetris. These players perform well in single games. In this paper, we developed a platform based on an open source project for game competitions among multiple players. We develop an artificial player employed learning by imitation, which is novel in Tetris. The imitation tasks of playing Tetris were mapped to a standard data classification problem. The experiments showed that the performance of the player can be significantly improved when our player acquires similar game skills as those of the imitated human. Our player can play Tetris in diverse ways by imitating different players, and has chances to defeat the best-known artificial player in the world. The framework supports incremental learning because the artificial player can find stronger players and imitate their skills.

2009

Dapeng Zhang, Cai Zhongjie, Chen Kefei and Bernhard Nebel.
A Game Controller Based on Multiple Sensors.
In In Proceedings of the Fifth International Conference on Advances in Computer Entertainment Tochnology (ACE 2009). 2009.
(Show abstract) (Hide abstract) (PDF)

A digital game is normally controlled by hand. Playing such a game requires only minimum hand movements. Rather than being easy and comfortable, this game controller is designed to be physically taxing for the players. It consists of several sensors, which makes a game more lively and forces the users to be more physically active. By using different mapping methods, one game can be played in several ways. The statistics gathered from the experiments show that even though the quality of control on the chosen fighting game is not as high as with a normal joystick, the developed controller is still preferred by most of the participants. It induces much more movement than a normal joystick.

2008

Armin Hornung and Dapeng Zhang.
On-Line Detection of Rule Violations in Table Soccer.
In Andreas R. Dengel, Karsten Berns, Thomas M. Breuel, Frank Bomarius and Thomas R. Roth-Berghofer (eds.), Proceedings of the 31st Annual German Conference on AI (KI 2008), pp. 217-224. Springer-Verlag 2008.
Poster.
(Show abstract) (Hide abstract) (PDF)

In table soccer, humans can not always thoroughly observe fast actions like rod spins and kicks. However, this is necessary in order to detect rule violations for example for tournament play. We describe an automatic system using sensors on a regular soccer table to detect rule violations in realtime. Naive Bayes is used for kick classication, the parameters are trained using supervised learning. In the on-line experiments, rule violations were detected at a higher rate than by the human players. The implementation proved its usefulness by being used by humans in real games and sets a basis for future research using probability models in table soccer.
Dapeng Zhang and Armin Hornung.
A Table Soccer Game Recorder.
In Video Proceedings of the IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS 2008). Nice, France 2008.
Digest and Video.
(Show abstract) (Hide abstract) (PDF) (PS.GZ)

Our table soccer robot can already challenge even professional human players. Next, the robot should play games by using human-like skills. As a foundation of this research, our table soccer game recorder can save and replay games played by humans. This video shows the construction and functionality of the recording system.

We use three types of sensors mounted on a regular game table. The movement of a game rod is measured by an optical distance sensor. Its turning is observed by a magnetic rotary encoder. Two laser measurement systems are synchronized to determine the position of the ball. The raw sensor data is smoothed by an approach using multi-model Kalman filter. We developed several software modules for the system. The modules provide a basis for the future research.
Dapeng Zhang.
Robot Plays Table-Soccer.
In Proc. of Dagstuhl Seminar (08372), Computer Science in Sport - Mission and Methods 2008. 2008.
Presentation (in .ppt) .
(Show abstract) (Hide abstract) (PDF)

Our research focuses on learning approaches with robot KiRo. KiRo is a table soccer robot which can challenge even advanced human players. Previously, we developed a method using learning by imitation, by which KiRo can automatically acquire the demonstrated actions. Recently, we constructed a game-recorder which collects data from the human-played games. The in-process work is about explaining the recorded data, which is to classify and to evaluate human's skills. A brief overview of the previous work is addressed, and the perspective is discussed.
Dapeng Zhang, Bernhard Nebel and Armin Hornung.
Switching Attention Learning - A Paradigm for Introspection and Incremental Learning.
In Proceedings of Fifth International Conference on Computational Intelligence, Robotics and Autonomous Systems (CIRAS 2008), pp. 99-104. Linz, Austria 2008.
(Show abstract) (Hide abstract) (PDF)

Humans improve their sport skills by eliminating one recognizable weakness at a time. Inspired by this observation, we define a learning paradigm in which different learners can be plugged together. An extra attention model is in charge of iterating over them and chosing which one to improve next. The paradigm is named Switching Attention Learning (SAL). The essential idea is that improving one model in the system generates more "improvement space" for the others. Using SAL, an application for tracking the game ball in a table soccer game-recorder is implemented. We developed several models and learners which work together in the SAL framework, producing satisfying results in the experiments. The related problems, advantages, and perspective of the switching attention learning are discussed in this paper.

2007

Dapeng Zhang and Bernhard Nebel.
Recording and Segmenting Table Soccer Games -- Initial Results.
In Proceedings of the 1st International Symposium on Skill Science 2007 (ISSS 2007), pp. 193-195. 2007.
(Show abstract) (Hide abstract) (PDF)

Robot KiRo can play one side of a table soccer game autonomously. Our recent research focuses on learning from and acting against human actions. Therefore recording and segmenting games played by humans are motivated. In this paper, the construction of a table soccer game recorder is sketched. An intuitive segmenting algorithm is implemented to explore the properties of the recorded data. A segmentation approach using Hidden Markov Models (HMMs) is proposed.
Dapeng Zhang and Bernhard Nebel.
Learning a Table Soccer Robot a New Action Sequence by Observing and Imitating.
In Proceedings of the Third Artificial Intelligence for Interactive Digital Entertainment Conference (AIIDE 2007), pp. 61-67. 2007.
(Show abstract) (Hide abstract) (PDF)

Star-Kick is a commercially available and fully automatic table soccer (foosball) robot, which plays table soccer games against human players on a competitive level. One of our research goals is to learn this table soccer robot skillful actions similar to a human player based on a moderate number of trials. Two independent learning algorithms are employed for learning a new lock and slide-kick action sequence by observing the performed actions and imitating the relative actions of a human player. The experiments with Star-Kick show that an effective action sequence can be learned in approximately 20 trials.

2005

Dapeng Zhang.
Action Selection and Action Control for Playing Table Soccer Using Markov Decision Processes.
Master's thesis, Albert-Ludwigs-Universität, Freiburg, Germany 2005.
(Show abstract) (Hide abstract) (PS.GZ) (PDF)

StarKick is a commercially available table soccer robot which challenges even advanced human players. However, the available set of actions for StarKick is limited and the way of selecting the actions is not flexible enough for incorporating more elaborate actions.

In the context of this thesis, new actions for stopping and dribbling the ball are developed. Stopping is achieved by locking the ball between the playing surface and a playing figure. Dribbling makes the ball rolling at a controllable speed within the reachable area of the playing figures of one rod. By these new actions, the ball can be deliberately passed and stopped.

To decide, which action should be taken in a given situation, an action selection scheme using Markov Decision Processes (MDPs) and reinforcement learning is developed. In order to reduce the state space, the basic actions are combined to more complex actions and the MDP is structured into four modules. Each module contains a set of states, and the actions that are applicable in these states. A simple reinforcement learning algorithm is implemented in the MDP framework. The transition probabilities are updated by counting. These updates are spread by policy iteration algorithm during a game.

A series of experiments are carried out in real table soccer games. These experiments show that the newly developed actions are robust, the MDP model works fine, and the reinforcement learning improves the performance of StarKick in a simplified game. The attempt of making the reinforcement learning work in the whole game seems too tedious to be finished in real games.

2004

Thilo Weigel, Dapeng Zhang, Klaus Rechert and Bernhard Nebel.
Adaptive Vision for Playing Table Soccer.
In S. Biundo, T. Frühwirth and G. Palm (eds.), KI 2004: Advances in Artificial Intelligence. Proceedings of the 27th Annual German Conference on Artificial Intelligence, pp. 424-438. Springer-Verlag 2004.
(Show abstract) (Hide abstract) (PDF)

For real time object recognition and tracking often color-based methods are used. While these methods are very ecient, they usually dependent heavily on lighting conditions. In this paper we present a robust and ecient vision system for the table soccer robot KiRo. By exploiting knowledge about invariant characteristics of the table soccer game, the system is able to adapt to changing lighting conditions dynamically and to detect relevant objects on the table within a few milliseconds. We give experimental evidence for the robustness and efficiency of our approach.