Action Selection for Game Play Agents Using Genetic Algorithms in Platform Game Computational Intelligence Competitions

(1)

Development Report:

Action Selection for Game Play Agents Using Genetic Algorithms in Platform Game Computational Intelligence Competitions

Ken Hasegawa, Narutoshi Tanaka, Ryuji Emoto, Yusuke Sugihara, Ardta Ngonphachanh, Junko Ichino, and Tomonori Hashiyama

Graduate School of Information Systems, The University of Electro-Communications 1-5-1 Chofugaoka, Chofu, Tokyo 182-8585, Japan

E-mail:{ken-hasegawa, tanaka, emoto, yusuke, ardta}@media.is.uec.ac.jp,{ichino, hashiyama}@is.uec.ac.jp [Received December 6, 2012; accepted January 28, 2013]

The application of computational intelligence (CI) and artificial intelligence (AI) to games has been attempted as a typical implementation of intelligent processing on computers. Intelligence in this sense is under- stood as the ability to search for the best solution efficiently among multiple options, specifically in AI playing board games such as chess. As the processing ability of computers increases, CI/AI systems are outperforming humans in finding potential solu- tions from a tremendous number of options within a short timeframe. These days, computer games are widely prevalent. CI/AI applications in computer games are focused on animating non-player characters (NPCs), whereas CI/AI applications in the sci- entific fields are focused on modeling intelligent human activities. The field of computer games faces many issues, such as dealing with dynamic environments that change quickly and processing images at higher resolutions and complexity. The use of computer games as a benchmark for CI/AI technologies has been attempted, and competitions involving various kinds of games have been held to encourage inno- vation in the field. In this paper, we describe a learning agent that participated in a platform game CI competition held in conjunction with Fuzzy System Sympo- sium (FSS 2012). The approach adopted in this paper is a basic method based on conventional methods.

The authors hope that this presentation of our development processes would encourage many researchers to participate in competitions and that it would contribute to progress in the field.

Keywords: platform games, genetic algorithm, game play agent, competitions

1. Introduction

The application of computational intelligence (CI) and artificial intelligence (AI) to the field of gaming has ac- complished much, defeating human champions [1] in board games, such as chess and shogi, and in card games,

such as poker, thereby contributing to the progress in CI/AI technologies. In recent years, the number of video game players increased with the spread of stationary game consoles and with the technological advancement of portable game terminals and smartphones. Therefore, the increase in CI/AI applications in computer games was also expected.

The success of computer games depends on numerous factors, including graphics quality, consistency in simu- lating physical behaviors, high quality of rendered sce- narios, ease-of-control, and intelligence of game characters. The development of graphics quality and physical engines is achieved through technological advancement in hardware. Recently, game developers differenti- ated themselves by constructing or introducing a unique world view [2]. CI/AI applications in computer games require several technologies, including those that recog- nize changes in a dynamic environment and then adapt accordingly or respond instantly [3]. Conventional CI/AI applications in video games have mainly been used to animate non-player characters (NPCs). Gamers quickly lose interest if an apparently high-definition NPC can only re- peat simple actions. An intelligent system is required to adapt to the sensitivity and skills of the players and to animate the NPC beyond static, predictable motions [2].

Yannakakis pointed out that game developers disregard academic CI/AI technologies, which they believe have not resolved practical issues; meanwhile, CI/AI researchers believe that the game industry does not use sophisticated technologies. Such a gap must be bridged, because the cooperation between businesses and universities is impor- tant in the future [4]. Furthermore, research studies on popular computer games are more likely to attract young researchers who would work on further developing CI/AI in the future.

In recent years, competitions for computer games are held in various international conferences. For example, competitions in the IEEE Conference on Computational Intelligence and Games (CIG 2012) focused on commer- cially successful video games or their clones, including a wide variety of first-person shooter (FPS), racing, real- time strategy (RTS), and platform games. We entered our learning agent [5] in the “Platform Game CI Competition

(2)

2012 [6],” which was held in Japan and was based on the IEEE CIG competitions.

In platform games, the characters are controlled by players by means of commands, such as ‘run’ or ‘jump,’

to move towards a goal in a configured stage, which could be 2-dimensional or 3-dimensional. In the competition we entered, the characters (agents) moved on a scrolling 2-dimensional plane with a variety of terrain (footing) configurations and headed toward a goal while beating enemies, overcoming obstacles, and retrieving items. Specifically, the competition was based on Infinite Mario Bros, a clone of the Super Mario Brothers games by Nintendo. The competition highlighted the “Learning Track,” a specific competition that focused on the quality of actions that agents performed after a limited learning period.

In this paper, the computer game competition is described in Section 2, and the agent developed by the authors is discussed in Section 3. Lastly, the results of the simulation and the summary are presented.

2. Game CI/AI Competition

Games have been considered as an ideal test bed for CI/AI studies. Applications that could defeat human champions in board games (e.g., chess or shogi) and in card games have been developed with positive results. Re- cent research studies focused on applying CI/AI to computer games [7]. In fact, IEEE Transaction on Computa- tional Intelligence and AI in Games [8] was launched in 2009. Therefore, future progress in this field is expected.

In this section, we outline the competitions for CI/AI applications on computer games, and we describe in detail the specific platform game CI competition, which we entered.

2.1. Game Competition

In the 2000’s, competitions for CI/AI game applications were held in IEEE international conferences and similar gatherings. For example, the Congress on Evo- lutionary Computation held competitions on repetitive prisoners’ dilemma games and on automatic control systems using video images obtained from radio-controlled cars. In 2005, a competition for video games was held in the “Computational Intelligence and Games” (CIG) conference. CIG 2012 included competitions using actual video games and their clones, including Unreal Tourna- ment 2004 (a first-person shooter game), StarCraft (a real- time strategy game), PacMan (an action game), and Mario Bros (a platform game). CI/AI challenges were established in each competition; for example, the participants competed by developing a bot that behaved like humans in first-person shooting games and by capturing information from images in PacMan. Three focused competitive tracks were established for Mario Bros: (1) the Game- play/Learning Track [9], where competitors develop an automatic game controller programmed to attain the highest scores; (2) the Level Generation Track [10, 11], where

Fig. 1. Example of platform game stage.

Fig. 2. The states of Mario.

competitors configure stages that are appropriate for the players’ skills; and (3) the Turning Test Track [12], where competitors produce controllers that emulate human behavior. In 2012, a platform game CI competition for the Learning Track using Infinite Mario Bros [6] was held in conjunction with the Fuzzy System Symposium (FSS 2012).

2.2. Platform Game CI Competition [6, 9]

2.2.1. Game Overview

The objective of the game is to reach a goal on the right- most end of the stages by controlling the character Mario at the center of the screen on a 2-dimensional plane, as shown inFig. 1. Mario may walk or run to the left or right, and jump. He could switch to any of three states:

Fire, Large, or Small, as shown inFig. 2, and he defeats enemies by shooting fireballs in the Fire state. Power-up items may appear, if blocks are hit from below. A power- up allows Mario to change to the Large or Fire state. The state also changes if Mario is damaged due to an attack by enemies. If Mario is damaged in the Small state, the game ends. In addition, falling into a hole also ends the game, regardless of the current state. Scores are calculated based on the elapsed time and the number of enemies defeated before reaching the goal, as well as the number of power- up items and coins obtained.

(3)

Fig. 3. Game grid.

2.2.2. API

The competition package [13] was encoded using the JAVA language and was based on Markus Persson’s Infi- nite Mario Bros, which is an open-source clone of Nin- tendo’s Super Mario Brothers. CI/AI programs controlling Mario were constructed using the following APIs in the Learning Track. The control cycle is 24 fps.

• Environment interface:

The environment was divided into grids as shown inFig. 3, and Mario could obtain the following environment information: (1) information on holes or ground shapes; (2) location information on enemies;

(3) Mario’s current state. Fig. 3 shows 5×5 grids as an example, but players could obtain environment information from 19×19 grids that encompass the whole screen.

• Agent interface:

Mario’s action was determined based on information obtained using the environment interface. One of five commands (right, left, beneath, jump, accelerate/fireball) could be chosen as Mario’s action, and multiple actions could be performed at the same time.

• Task interface:

The competition was restricted by predetermined parameters, such as the difficulty level and the random seed values, or by calculation methods for scores and evaluation functions for generated Mario agents.

2.2.3. Learning Track

In the Learning Track competition, Mario agents learned within the limits of the task interface parameters, such as the difficulty level, the random seed values, and the learning times, which were predetermined by the or- ganizers. Competitors implemented Mario agents using CI/AI technology frameworks. In the actual competition, trained agents competed for scores in competition stages that were different from those in the learning phase. The winning source code will be open-sourced to ensure fair- ness.

3. Agent Design

The parameters for the competition were determined as follows:

• Learning iterations: a total of 10,000 times

• Difficulty level: 0, 3, 5, 10

• Random number seed: 0

3.1. Game Analysis

Before designing learning agents for platform game, the authors first played the game a number of times, then observed and analyzed the norm in terms of difficulty levels and human control methods.

3.2. Difficulty Levels

The difficulty level of the games could be determined using the task interface described in the previous section about the API. Findings obtained by actually playing the game with the difficulty level from 0 to 10 are presented below.

• Hole width:

No holes appear at Level 0. The width of holes is 3 boxes wide at Level 4 or lower. The width can be 10 boxes wide at Level 5 or higher. The width is always 10 boxes wide at Level 8 or higher.

• Type of enemies:

Enemies walk on the surface at Level 0. At Levels 1 and 2, enemies appear that fly with wings or that only stop when stepped on. At Level 3 or higher, some enemies cannot be defeated even with fireballs.

• Number of enemies:

The number of enemies is limited to the number of the level+1.

These findings can be confirmed by the definition in the LevelGenerator class included in the package.

3.3. Specifications of the Learning Agent

The following design policy was determined after or- ganizing the above-mentioned findings and after several brainstorming sessions.

The basic actions of the agent must be meta-level tasks such as escape or get items, instead of direct commands such as right, left, beneath, and jump.

Low difficulty stages with fewer enemies and sim- pler ground shapes and obstacle arrangements could be cleared relatively easily, if commands (right, left, beneath, jump) were appropriately performed. However, in high difficulty stages, players could deal with more enemies by attacking or escaping from them. In order to jump over 5 boxes, the character must accelerate prior to approach or jump several times using floating islands above the holes.

(4)

We decided that direct control commands, such as right, left, beneath, and jump, are not sufficient to pass high difficulty stages during the prescribed learning times. Meta- level actions, such as attack and escape, were hard-coded in the implementation, and the appropriate action was selected based on intelligence acquired by learning.

We determined that the agent should be able to choose among the following six meta-level actions. The actions that deal with enemies are escape and attack.

1. Jump over holes.

2. Escape from enemies.

3. Attack enemies.

4. Get items.

5. Move forward.

6. Do nothing.

Falling into a hole ends the game, regardless of Mario’s current state; therefore, jumping over holes is judged to be the top priority action and will always be selected when Mario encounters a hole. A system for incorporating fea- tures, such as hierarchical module learning and learning meta-level actions, will be studied and incorporated in the future.

3.4. Learning Actions Using Genetic Algorithms We implemented the six meta-level actions listed in the previous section. Genetic algorithm (GA) was used to learn which actions to select in response to environmental information [14]. The implemented genetic algorithm is described below.

3.4.1. Environment Interface

The following 11 bits were used to store information obtained through the environment interface.

• Whether there are enemies near forward, near backward, or far forward: 3 bits.

• Whether there are holes backward, near forward, forward in the middle distance, or far forward: 4 bits.

• Whether Mario is stopped, running forward, walking forward, or moving backward: 2 bits.

• Whether Mario is tall or short: 1 bit.

• Whether an item blocks the path: 1 bit.

Environment information is 00000000100 in the case ofFig. 4. Only one bit is set to 1, which indicates that Mario is moving forward. Digits are set to 0, if the corre- sponding conditions are not satisfied.

3.4.2. Action Selection

As shown inFig. 4, output information is designed to select from the six meta-level tasks discussed in the previous section. Jumping over holes is always selected because this action is performed first, when holes are en- countered. Therefore, this action is excluded from the

Input 0 0 0 0 0 0 0 0 1 0 0

Enemy Hole

NearFar

Speed Pos

Near

Item

15

Chromosome

00000000100⁽²⁾=4⁽¹⁰⁾

[0] [1] [2] [3] [4] [5] [4095]

15⁽¹⁰⁾=01111⁽²⁾

Output

＊ ⁰ ¹ ¹ ¹ ¹ * : Default

1 : On 0 : Off Jump (over hole)

Escape Attack

Get Item Forward

Do Nothing

Priority High

Low

Fig. 4. Coding to chromosome.

chromosome output of GA and is indicated as ‘*’ in Fig. 4. Five bits are assigned to the remaining five actions and the priorities are set as shown; multiple actions can be selected at the same time. In the output information in Fig. 4, all actions besides “escape from enemies” are selected. In this case, the agent moves forward to get items and to attack enemies. The agent must do something and move forward at the same time. If the agent is unable to select multiple actions on its own, an action with a high priority is performed in addition to moving forward.

3.4.3. Coding to Chromosome

As shown inFig. 4, the chromosome in GA was designed to indicate what action is selected in response to given environmental information. The correlation is shown below, using Fig. 4 as an example. The environment information represented by binary numbers of 11 bits is converted to a decimal number (4 in the case of Fig. 4). This shows the genetic locus. Specifically, the chromosome sequence has a zero-based index, and the element at index 5 expresses the action of the agent in this environment. The action is represented as a decimal number from 0 to 31 (15 in the case ofFig. 4) and then converted to 5-bit binary numbers. Each of the 5 bits corresponds to actions (except jumping over holes) in the order of their priority. Actions are not performed if the value is 0 and performed if the value is 1.

4. Competition

4.1. Setting Parameters of Genetic Algorithms We participated in the competition after implementing the learning agent using GA as described in the previous

(5)

Table 1. Score after learning using GA.

total Lv0 Lv3 Lv5 Lv10

1 8198 4096 2231 1490 381

2 7830 4096 2929 250 555

3 7725 4096 2790 334 505

4 7715 4096 2236 865 518

5 7706 4096 2236 845 529

6 7706 4096 2236 845 529

7 7698 4096 2379 395 828

8 7697 4096 2236 930 435

9 7683 4096 2231 979 377

10 7675 4096 2053 691 835

before learning 6010 4096 716 788 410

section. General operations were used for GA calcula- tions: mutation by bit inversion and one-point crossover.

Elitist strategy and roulette wheel selection were used to selection of chromosomes. When learning with GA is performed, mutation rate and crossover rate should be set. The total learning iterations are predetermined to be 10,000 times for the competition. In other words, the number of individuals multiplied by the number of gener- ations must equal 10,000; therefore, an appropriate number for individuals must be determined. In order to set these learning parameters, the simulation for setting parameters was conducted with 8 patterns for the number of individuals, 20 patterns for every 5% of mutation rate and crossover rate, which is a total of 8×20×20 patterns.

Based on the simulation results, we determined the ideal number of individuals to be 100, the mutation rate to be 10%, and the crossover rate to be 50%.

4.2. Learning Results

The learning results of agents differed, because the learning times were set to as few as 10,000 times. GA learning was conducted as time allowed using the learning parameters described in the previous section. Specifically, the simulation was conducted by frequently updating the 10 highest-scoring patterns after learning 10,000 times.

Table 1shows the performance results before and after learning. The results of the 10 highest-scoring attempts at the start of the competition are shown as the post-learning results.

Level 0 (Lv0) was cleared in all cases before and after learning. At Level 3 (Lv3), the score increased from 716 to more than 2000 due to learning. At Levels 5 and 10 (Lv5 and Lv10), it was difficult to conclude the effect of learning, because we could not determine any par- ticular tendency, and a trade-off relationship appeared to emerge. In short, cases with a high score at Lv5 showed a low score at Lv10, and vice versa. As analyzed in Sec- tion 3, the width of holes and the number of enemies increased at higher difficulty levels. A qualitative charac- teristic was difficult to extract from control rules obtained

Table 2. Main competition results.

Team Name Score Methods UEC-IS Team [5] 13,511 GA Mie Univ. Team [15] 11,896 Q learning

Nitech Team [16] 11,260 GA

from the learning results of GA, but the present implementation does not necessarily respond well at difficulty levels higher than Lv5, as can be seen in plays using the learning results. The case with the highest point of 1490 at Lv5 was submitted to the main competition.

4.3. Regulations and Results of Main Competition The following regulations are specified in the main competition.

• Score: Provide the total distance covered by the Mario agents.

• Tie-break: Consider the number of gained coins and the number of enemies defeated, and remained time when the scores are the same.

• Difficulty level: Include a total of 10 stages with 2 stages each for Levels 0, 1, 4, 7, 10.

Each team brought agent files to the competition after 10,000 learning attempts based on the tasks described in Section 3. Under the above regulations, agents performed in the competition site, and scores were calculated.

Three teams participated in the competition [5, 15, 16].

The results of the competition are shown inTable 2along with the methods used for agents by each team. The authors submitted the learned Mario agent discussed in the previous section and won with a score of 13,511 points.

Refer to the Web site [17] for more information about the competition.

5. Conclusion and Future Issues

We introduced a learning agent for the platform game CI competition held in conjunction with FSS 2012. The proposed agent adopts meta-level actions, such as escaping from and attacking enemies, as basic elements instead of the direct-control commands of right, left, beneath, and jump. A genetic algorithm was used for the method that selects meta-level actions in response to environments.

The effect of learning was demonstrated by using the proposed system. However, whether the proposed method is adequate for problem resolution is unknown and an open question. For example, future CI/AI studies need to define meta-level actions and to determine actions by learning even without prior definitions.

In the competition, it was pointed out from the audience that a few reference materials in the introductory package

(6)

were written in Japanese. We prepared a simple introduction manual based on their experience [18]. The authors wish to contribute to the progress in the field through competitions in the future by broadening the application range of CI/AI technologies.

References:

[1] M. Campbell, A. J. Hoane Jr., and F. Hsu, “Deep Blue,” Artificial Intelligence, Vol.134, Issues 1-2, pp. 57-83, 2002.

DOI: 10.1016S0004-3702(01)00129-1

[2] T. Hartley and Q. Mehdi, “Online action adaptation in interactive computer games,” ACM Computers in Entertainment, Vol.7, No.2, Article No.28, p. 31, 2009. DOI: 10.11451541895.1541908 [3] S. Karakovskiy and J. Togelius, “The Mario AI Benchmark and

Competitions,” IEEE Trans. on Computational Intelligence and AI in Games, Vol.4, No.1, pp. 55-67, 2012.

[4] G. N. Yannakakis, “Game AI revisited,” Proc. of the 9th Conf. on Computing Frontiers, pp. 285-292, 2012.

DOI: 10.1145/2212908.2212954

[5] N. Tanaka, R. Emoto, K. Hasegawa, Y. Sugihara, A.

Ngonphachanh, J. Ichino, and T. Hashiyama, “A game play agent learns to play as if it is controlled by human player,” Proc. of 28th Fuzzy System Symposium, pp. 292-293, 2012 (in Japanese).

[6] T. Takano, T. Ichimura, M. Kanoh, and M. Koshino, “The Platform Game CI Competition 2012 in Japan,” Proc. of 28th Fuzzy System Symposium, pp. 516-517, 2012 (in Japanese).

[7] J. E. Laird, “Research in human-level AI using computer games,”

Communications of the ACM, Vol.3, No.8, pp. 32-35, 2002.

[8] S. M. Lucas, “Computational Intelligence and AI in Games: A New IEEE Transactions,” IEEE Trans. on Computational Intelli- gence and AI in Games, Vol.1, No.1, pp. 1-3, 2009.

[9] J. Togelius, S. Karakovskiy, and R. Baumgarten, “The 2009 Mar- ioAI competition,” Proc. IEEE Congress on Evolutionary Compu- tation, 2010. DOI: 10.1109/CEC.2010.5586133

[10] J. Togelius, G. N. Yannakakis, B. Weber, T. Shimizu, T. Hashiyama et al., “The 2010 Mario AI Championship: Level Generation Track,” IEEE Trans. on Computational Intelligence and AI in Games, Vol.3, No.4, pp. 332-337, 2011.

[11] T. Shimizu, T. Hashiyama, T. Esaki, J. Ichino, and S. Tano, “Online Level Generation using User Profiles to Keep the user in the Flow Channel,” Proc. of 27th Fuzzy System Symposium, pp. 557-562, 2011 (in Japanese).

[12] J. Togelius, G. N. Yannakakis, S. Karakovskiy, and N. Shaker, “As- sessing Believability,” Ch.9 in Believable Bots, P. Hingston (Ed.), pp. 215-230, Springer 2012. DOI:10.1007/978-3-642-32323-2 9 [13] Platform Game CI Testbed package.

http://www.marioai.org/gameplay-track/getting-started

[14] T. Takano and M. Kanoh, “Platform Game AI Tutorial,” Proc. of 33rd Tokai Fuzzy Symposium (Gama-ken), 2012 (in Japanese).

[15] S. Shinohara, T. Takano, H. Takase, H. Kawanaka, and S. Tsuruoka,

“A discussion on planning method in platform game AI – A study of situation recognition for utilizing learning data –,” Proc. of 28th Fuzzy System Symposium, pp. 705-708, 2012 (in Japanese).

[16] M. Maeda, H. Kanii, T. Nakamura, M. Kanoh, and K. Yamada, “Ob- taining Action Rules for Game Agents with Evolutional Computa- tion,” Proc. of 28th Fuzzy System Symposium, pp. 713-714, 2012 (in Japanese).

[17] Result of Platform Game CI Competition 2012 in Japan.

http://sns.j-soft.org/community/82/reference/22532 [18] Mario AI Manual (in Japanese).

http://www.media.is.uec.ac.jp/medialab-wp/imlab/resources/

Name:

Ken Hasegawa

Affiliation:

Graduate Student, Department of Human Me- dia Systems, Graduate School of Informa- tion Systems, The University of Electro- Communications

Address:

1-5-1 Chofugaoka, Chofu, Tokyo 182-8585, Japan Brief Biographical History:

2012 B.S. in Information Technology, The University of Electro-Communications

2012- Graduate School of Information Systems, The University of Electro-Communications

Main Works:

•“A game play agent learns to play as if it is controlled by human player,”

Proc. of 28th Fuzzy System Symposium, pp. 292-293, 2012 (in Japanese).

Membership in Academic Societies:

•Japan Society for Fuzzy Theory and Intelligent Informatics (SOFT)

Name:

Narutoshi Tanaka

Affiliation:

Address:

2008-2012 B.S., The University of Electro-Communications 2012- Graduate School of Information Systems, The University of Electro-Communications

Main Works:

•research of learning AI in computer games Membership in Academic Societies:

Name:

Ryuji Emoto

Affiliation:

Graduate Student, Master’s Course, The Univer- sity of Electro-Communications

Address:

2012 B.S, Tokyo Metropolitan College of Industrial Technology 2012- Graduate School of Information Systems, The University of Electro-Communications

Main Works:

(7)

Name:

Yusuke Sugihara

Affiliation:

Address:

2012 B.S, Polytechnic University of Japan

Main Works:

Name:

Ardta Ngonphachanh

Affiliation:

Graduate Student, Graduate School of Infor- mation Systems, The University of Electro- Communications

Address:

2012 B.S, Polytechnic University of Japan

Main Works:

Name:

Junko Ichino

Affiliation:

Graduate School of Information Systems, The University of Electro-Communications

Address:

1996-1998 M.S. in Engineering, University of Electro-Communications 1998-2001 System Engineer, Dai Nippon Printing Co., Ltd.

2001-2006 System Engineer and Project Reader, TIS Inc.

2003-2006 Researcher, National Institute of Information and Communications Technology (NICT)

2004-2007 Ph.D. in Engineering, Kobe University

2010 Visiting Researcher, The Interactions Laboratory, Department of Computer Science, University of Calgary, Canada

2007- Assistant Professor, University of Electro-Communications Main Works:

•“Effects of the display angle in museums on user’s cognition, behavior, and emotions,” Proc. CHI 2013, 2013.

•“Vuzik: A Painting Graphic Score Interface for Composing and Control of Sound Generation,” Proc. ICMC 2012, pp. 579-583, 2012.

•“Discriminating Divergent/Convergent Phases of Meeting Using Non-Verbal Speech Patterns,” Proc. ECSCW2011, pp. 153-172, 2011.

•The Association for Computing Machinery (ACM)

•Information Processing Society of Japan (IPSJ)

•Human Interface Society

Name:

Tomonori Hashiyama

Affiliation:

Associate Professor, Graduate School of In- formation Systems, The University of Electro- Communications

Address:

1996-1997 Assistant Professor, Nagoya University

1997-1999 Senior Researcher, Nagoya Industrial Science Research Institute

2000-2003 Associate Professor, Nagoya City University

2001-2002 Visiting Scholar, University of California, San Diego (UCSD) 2003- Associate Professor, The University of Electro-Communications Main Works:

•“The 2010 Mario AI Championship: Level Generation Track,” IEEE Trans. on Computational Intelligence and AI in Games, Vol.7, No.4, pp. 332-347, 2011.

•The Institute of Electrical and Electronics Engineers (IEEE)

•The Association for Computing Machinery (ACM)