IEEE CIG 2009 Conference Report [Conference Report]

0 downloads 0 Views 1MB Size Report
Pier Luca Lanzi,. Politecnico di Milano, Italy ... charmingly old buildings with gardens surrounded by .... Player Experience in Super Mario Bros” by Georgios ...
Conference Report

Pier Luca Lanzi, Politecnico di Milano, Italy David Robles Contreras, University of Essex, U.K., and Luigi Cardamone, Politecnico di Milano, Italy

IEEE CIG 2009 Conference Report

I

n September 2009, the weather in Milan was warm and sunny. It was the perfect setup for the 2009 IEEE Symposium on Computational Intelligence and Games (CIG-2009) held from September 7–10 at the Politecnico di Milano (http://www.polimi.it). The main campus, founded in 1837, has charmingly old buildings with gardens surrounded by porches to keep you dry during the rainy season and cool in the hot summer days. The symposium was hosted in the main hall of the south building, in the heart of the campus, just a few steps away from the beautiful Piazza Leonardo da Vinci, where students enjoy the sun during lunch breaks when the weather allows. CIG-2009 was the fifth in a series of exciting and vibrant events that started with CIG-2005 at the University of Essex. This year, 76 papers from all over the world were submitted and 52 were accepted for presentation and publication in the proceedings. The presentations covered a wide range of topics including Evolutionary Computation, Neural Networks, Machine Learning, etc. The program also included four keynotes, two tutorials and five competitions. CIG-2009 was organized as a rather relaxed event with late starts, with the first talk at 9am, and late-morning keynotes at 11am. Each keynote was followed by two pre sentations selected from the best eight accepted papers.

Digital Object Identifier 10.1109/MCI.2010.936324

20

Tutorials

The conference opened with the tutorial session. In the first tutorial, Practical Issues in Evolving Neural Network Controllers for Video Game Agents, Kenneth O. Stanley from University of Central Florida discussed how evolutionary algorithms can be effectively combined in video games. In particular, he demonstrated how to evolve a robot army in the game NERO, by tuning the artificial brains of team members to perform challenging tasks, such as learning to surround walls to get to a goal, to avoid damage, and many other tasks. In the second tutorial, Modeling and Simulation Framework of Real Urban and Board Games to Train Players, James M.Vaccaro from Lockheed Martin, focused on computer-simulated urban search and rescue operation and RISK, a multi-player non-cooperative stochastic game problem. In the tutorial, he discussed a modeling approach for generating urban terrain models for computer-simulated urban search and rescue operations; he also presented a modeling approach for implementing the game RISK and generating autonomous players.

speakers from industry (Microsoft Research Cambridge and Milestone) and two speakers from university (Reykjavik University and University of California at Santa Cruz). Stefano Lecchi (Milestone) gave the first keynote, titled “Artificial Intelligence in Games” regarding the behavior of non-player characters in racing games from his experiences in the game industry. He stated that in commercial racing games, the challenges for the development of a successful artificial intelligence (AI) translate into the programming of an AI which can adapt to the driving styles and to the driving capabilities of human players so as to improve their gaming experience. He also pointed out that in racing games, the behavior of non-player characters should be plausible, challeng ing throughout the game, adaptive, and that

Keynotes

CIG-2009 had four keynote sessions, each one comprising a 45-minute keynote and two presentations chosen from the eight best papers accepted for publications. The keynote program had a nice 50-50 balance between industry and academia with two

IEEE COMPUTATIONAL INTELLIGENCE MAGAZINE | MAY 2010

The day before, the general and the local chair bring all the material to the conference site.

it should also lead to realistic group behaviors. The keynote was followed by the presentation “Temporal Difference Learning with Interpolated Table Value Functions” by Simon Lucas and the presentation “Improving Temporal Difference Game Agent Control Using a Dynamic Exploration Rate During Control Learning” by Leo Galway. Yngvi Björnsson (Reykjavik University, Iceland) gave a keynote about the state-of-the-art of General Game Afternoon coffee break under the Politecnico’s porch. Playing (GGP) systems. The aim of GGP is to create intelligent agents that Video Game” by Kenneth O. Stanley can automatically learn how to play a and “A Simple Tree Search Method for wide variety of different games at an Playing Ms. Pac-Man” by David Robexpert level without any game-specific les. knowledge. He discussed his SimulaDavid Stern (Microsoft Research tion-Based General Game Player, Cambridge, UK) gave an exciting keyCADIAPLAYER, which proved its note about the Computation Intellieffectiveness by winning the 2007 and gence techniques deployed in some 2008 AAAI GGP competitions using Microsoft commercial games. Initially, Monte Carlo simulations for its move he focused on the application of reindecisions. The keynote also included difforcement learning to fighting games. ferent design models, and discussed Then, he discussed how imitation learnsome open research challenges that ing has been applied in Forza MotorGGP poses for artificial intelligence subsport where the player can train a disciplines, such as knowledge represenDrivatar (based on his own driving style) tation, agent-based reasoning, heuristic which can replace the player in less search, computational intelligence, and interesting tracks. The keynote ended machine learning. The keynote was folwith a presentation of skill matching lowed by the presentations “Modeling techniques which can increase the chalPlayer Experience in Super Mario Bros” lenge in on-line multi-player matches of by Georgios Yannakakis and “A Data Halo 3. The keynote was followed by Mining Approach to Strategy Predicthe presentation “Optimized Sensorytion” by Michael Mateas. motor Couplings plus Strategy ExtenIn the third keynote, titled “AI Isn’t sions for the TORCS Car Racing Just for Players: AI-based Authoring Challenge” by Martin V. Butz and Tools”, Michael Mateas (University of “Evolving Multi-modal Behavior in California at Santa Cruz, USA) disNPCs” by Jacob Schrum. cussed the problems related to the game authoring process. He suggested that game AI research should not only focus on improving the player experience (as done so far) but also on helping authors in creating the richness and nuances that make games compelling. Accordingly, he overviewed a number of authoring support tool projects taking place in the Expressive Intelligence Studio at UC Santa Cruz. This keynote was followed by the presentation “Evolving Simon Lucas introducing the Ms Pac-Man Content in the Galactic Arms Race competition.

Presentations

Overall, the symposium had one poster session, held during the reception on the very first conference day, and nine presentation sessions, each one devoted to a specific topic, including, game theory, board and card games, modeling, planning, evolution, Ms Pac-Man, Super Mario, simulated car racing and first person shooters. As in past years, several interesting works have been presented covering all the aspects of the Computational Intelligence and Games field. On one hand, robust and mature techniques have been applied to last generation computer games like Unreal Tournament, Quake, TORCS and several Real Time Strategy games, showing that Computational Intelligence has great potentials also from the point of view of the game industry. On the other hand, more theoretical works focused on the technique itself, and applied improved or even new approaches to standard benchmarks like Pac-Man or well-known board games that provide a more accurate comparison with previous works. Many works focused on the evolution of non-player characters but also new and interesting directions including imitation learning, game testing, measuring player experience, player modeling and automatic game content generation. As usual, the proceedings are freely available on-line from the CIG-2009 website (http://www.ieee-cig.org/cig2009/Proceedings/) while the papers from the previous events are available at the webpage of the CIG conferences (http://www.ieee-cig.org/). Competitions

The conference included five exciting competitions which gathered great interest in the community and received many submissions. The Defcon competition was carried out for the first time during a CIG event. Defcon is a strategic game like RISK and the aim of the competition is to develop the best Defcon bot. Each bot played 30 matches against each other in a series of one-on-one

MAY 2010 | IEEE COMPUTATIONAL INTELLIGENCE MAGAZINE

21

In a hidden room, human players play against the judges.

The Mario AI Competition (http:// julian.togelius.com/mariocompetition2009/) consisted of controlling the Mario character through a series of levels.The competition received 15 submissions and Robin Baumgarten was the winner with an approach based on Aw. The other submissions showed a variety of techniques: evolutionary computation, modular architectures, rule-based architectures and neural networks. Interestingly, some of the submissions of the different competitions resulted in 6 high quality papers showing that competitions are powerful means to push in some promising research directions. And the Winner Is...

matches with bots playing in each of the possible starting territory configurations. The game ran on a limited-information mode (i.e., bots could not see units hidden by the fog of war) and for each match, the resulting score was recorded for each player. In the end, the player with the highest cumulative score won. The 2K Botprize was carried out for the second time in a CIG event. The aim of the contest was to test whether a computer game playing bot could play like a human in a First Person Shooter. In the contest, bots tried to convince a panel of expert judges that they were actually human players. None of the bots were able to fool enough judges to take the main prize (7000 Australian Dollars!), but all the bots fooled at least one of the judges. The competition took place during lunch time and it was hosted in the Educafé, a small cozy space in the heart of the Politecnico’s North building, and in a computer room just below the South building rooftop. Pizza and softdrinks were served during the event! The most human-like bot was sqlitebot by Jeremy Cothran and the joint runners-up were anubot from Chris Pelling and ICE-2009 from the team from Ritsumeikan University, Japan. The Ms. Pac-Man competition consisted of developing an agent controller to play Ms. Pac-Man. The agent’s input was the streaming real-time video output from the original game (i.e. the agent did not have access to any details of the software platform). The aim of the

22

Ms. Pac-Man agent was to score as many points as possible; the winner was the agent that achieved the highest score over ten runs each. There were four functioning entries submitted by the deadline, and the winning team was ICE Pambush 3, by Hiroshi Matsumoto, Takashi Ashida, Yuta Ozasa,Takashi Maruyama, and Ruck Thawonmas (Ritsumeikan University, Japan), with a new screen capture software agent Ms Pac-Man’s world record of 30,010 points! This run was performed live during the Ms. Pac-Man competition session at the conference, and was exciting to watch, with the entry surviving many seemingly impossible situations. The simulated car racing competition of CIG-2009 was the final event of the 2009 Simulated Car Racing Championship (http://cig.dei.polimi. it/?page_id=81), an event joining the three competitions held at CEC-2009, GECCO-2009, and CIG-2009. The aim of the competition was to develop a controller capable of racing against other opponents in three unknown tracks. This leg of the championship was won by Martin V. Butz and Thies D. Lönnecker while Enrique Onieva and David E. Pelta won the championship. The results demonstrated that, in the last year, the quality of the submissions had strongly improved: the controllers presented several complex behaviors (driving, overtaking, recovery, etc.) and also techniques to dynamically optimize some parameters during the race.

IEEE COMPUTATIONAL INTELLIGENCE MAGAZINE | MAY 2010

The committee assigned two awards, one for the best paper and one for the best student paper. The best paper award was assigned to Erin Hastings, Ratan Guha and Kenneth Stanley for their paper “Evolving Content in the Galactic Arms Race Video Game” (http://www.ieee-cig.org/cig-2009/ Proceedings/proceedings/papers/ cig2009_033e.pdf). The best student paper award went to Jacob Schrum for the paper “Evolving Multi-modal Behavior in NPCs” (http://www.ieeecig.org/cig-2009/Proceedings/proceedings/papers/cig2009_044e.pdf) he co-authored with Risto Miikkulainen. Supporters

CIG-2009 was sponsored by the IEEE Computational Intelligence Society (http://www.ieee-cis.org) and by the Politecnico di Milano. It also received donations from the IEEE Italian Chapter, IT-University of Copenhagen, and four companies: Microsoft Research Cambridge (http://research.microsoft.com/ en-us/labs/cambridge/), Milestone (http://www.milestone.it), NVidia (http://www.nvidia.com) and 2K Australia (http://www.2kaustralia.com/). Conclusions

CIG-2009 was a great experience and we look forward to the next edition, CIG-2010 (ht t p ://g a me.it u.d k / cig2010/), that will take place in Copenhagen, August 18–21, 2010.