Ms Pac-Man Competition
(screen capture mode)
IEEE CIG 2011 Results
Simon M. Lucas

Entries

Congratulations go to Nozomu Ikehata and Takeshi Ito of the University of Electro-Communications, Tokyo.  Their MCTS-based entry set a new high score of 36,280 for this competition, beating the previous record of ICE Pambush 3's of 30,010 achieved at IEEE CIG 2009.

There were five functioning entries as listed below.  While there is still room for improvement, the winning entry displayed some extremely skilful manoeuvres, escaping near death situations on so many occasions, and completing difficult mazes with many pills remaining even after all the power pills had been consumed.  Read the paper [1] and find out how they did it!

There are plans to run the competition again for IEEE CIG 2012 - join the cigames Google group for updates on this.

Entry / Author(s) / [reference] Affiliation High Score
Nozomu Ikehata and Takeshi Ito [1] The University of Electro-Communications, Tokyo 36,280
ICE Pambush 5 (Ruck)

Takeru Miyama, Asami Yamada, Yuki Okunishi, Takashi Ashida, and Ruck Thawonmas

[description]

Ritsumeikan University, Kyoto 27,240

Kuan-Wei Chen, Chu-Ming Wang, and
Tsung-Che Chiang
[description]

National Taiwan Normal University, Taiwan 20,300
SungUk Jeong and Kyung-Joong Kim
[description]

Sejong University, Seoul

19,900
Bruce Tong
[2][3]
City University of Hong Kong 13,700


Results

The full results are shown below.

CIG 2011 Results 

 

The winning entry is Nozomu, the one that achieved the highest score given thirteen runs each (ten prior to the conference and three live runs at the Ms Pac-Man competition session).  The averages are also shown for interest, and this time the winning entry also had the highest average.

Some obvious weaknesses of all the entries (copied from last year's competition - these points still apply, though to a lesser degree in the case of the winner):

  • they all have a fairly greedy short term strategy, and in many cases single food pills are left behind in the pursuit of more direct rewards, which makes life difficult towards the end of the level when these must be consumed in the absence of the relative security provided by the energiser pills.

  • they are not fully competent: they sometimes make unforced errors, and die by failing to make an obvious turn for example.  In some cases this appears to be due to screen parsing and/or control software failing to send the correct key response in sufficient time

We look forward to improved entries for IEEE CIG 2012, and there is a very good chance that we'll see one of these breaking the 50k barrier. 

Links

When you have a technique that works well, you may want to write it up as a paper for the IEEE Transactions on Computational Intelligence and AI in Games.

References

We encourage entrants to try evolutionary and machine learning methods.  The papers below might provide a useful starting point.

Also see the Pac Man papers in IEEE CIG 2008 and in IEEE CIG 2009, 2010, and 2011.

[1] Nozomu Ikehata and Takeshi Ito, Monte-Carlo Tree Search In Ms. Pac-Man, Proceedings of IEEE Conference on Computational Intelligence and Games 2011, Seoul, Korea.

[2] B. K. B. Tong, C. M. Ma, and C. W. Sung, A Monte-Carlo Approach for the Endgame of Ms. Pac-Man, Proceedings of IEEE Conference on Computational Intelligence and Games 2011, Seoul, Korea.

[3] B. K. B. Tong and C. W. Sung, A Monte-Carlo Approach for Ghost Avoidance in the Ms. Pac-Man Game, Proceedings of International IEEE Consumer Electronics Society's Games Innovations Conference (ICE-GIC), pp. 1-8, Hong Kong, Dec., 2010.

 

Also see: Simon M. Lucas, Evolving a Neural Network Location Evaluator to Play Ms. Pac-Man, Proceedings of IEEE Symposium on Computational Intelligence and Games 2005.

 

end of page