Friday 11 April

We arrived in the afternoon. We shared the room with the Soccer simulation, what meant that we had to be careful with the fileserver / network. The resources were excellent: GHz computers equipped with Suse 8.1 Linux.

The Freiburg team had already installed the kernel (version 4.2beta), and was assisting the Arian team with getting their code compiled for this environment. After two hours we also had our code running (we had developed our code with gcc 2.96).

In a teamleader meeting we decided use the Saturday to explore a number of maps and settings, and to decide on the end of the day what the most challenging final would be. At that moment we assumed half an hour running time per scenario, so for four teams we proposed the following schedule:

10:00 Virtual City (sf1 from Fukuoka 2002)

12:00 Kobe (final from Fukuoka 2002)

14:00 Freiburg

16:00 Amsterdam

18:00 Foligno

Freiburg was an experimental map developed by a student of the Freiburg team.

Later we found out that he had spent 20 hours to create this map, which made the creation of an Amsterdam-map during the weekend impossible.

Saterday 12 April

The Arian team had been working hard to create binaries for the GO environment.

Initial runs showed that they were loosing cycles and messages, which they couldn't solve remotely.

Also Freiburg had a lot of performance problems. At the end they found out that their simulation only worked when the kernel and agent SW ran on the same machine, otherwise they were loosing messages. With our SW it was vice versa. We needed at least two machines, otherwise we were losing cycles.

After a while we realized that the latest version of the kernel had problems with the settings of the 1st semi-final of Fukuoka 2002. The first game we played with the standard settings (delivered with the kernel) of the Virtual City, although the number of agents there are not enough to seriously rescue the City and its Civilians. Later we found that the kernel was able to run with the settings of the 2nd semi-final (sf2). The Virtual City run took 9 hours to complete.

UvA Rescue C2003	ResQ Freiburg

The next game we played with the Kobe map. In principal we used the settings of the final ofá Fukuoka 2002. The Freiburg team had made a slight modification: they changed the location of one FirePoint to make it more challenging. We were at that time far behind schedule, because the kernel becomes terrible slow when the fire spreads. The running times was in the order of 2 hours.

ResQ Freiburg	UvA Rescue C2003

The kernel had problems with the Freiburg map. There are some blank spaces in the center of the city, but we didn't had time to analyze if we had maybe an old version of the map.

We spent a lot of time to modify the Foligno map. Scaling inside JGISedit doesn't scale the properties of the buildings, so we tried to make a program that reads in the building.bin and scale its properties. Unfortunately, that tool was not ready on Saturday.

In the mean time we had manually modified the building properties in the South/East of the city. By restricting the firepoints to the South/East we could simulate a fire that burned down the modified buildings (1/3 of the City). We decided to use this map for the Final.

Sunday 13 april

10 minutes before the final the automatic scaling program was ready. No time to test, so we stayed with the manual scaled map. The Foligno map is quite challenging, because in principal there are not enough agents to seriously save the city. Yet, some of the reported problems (distances too large to travel) were not really a problem. The difficulty with the larger distances is more related with perception, and forces the agents to actively explore the environment.

With the ambulances this situation was extreme, the ambulances didn't hear the Civilians plies for help inside buildings, even when they were in front of the building.

This was not really a problem of distance, because for most buildings to distance between center and road is less than the hearing threshold of 30m.

ResQ Freiburg	UvA Rescue C2003

Because of performance problems we decided for a second run. In the mean time the Freiburg team had changed the behavior of their ambulances. In the second run their ambulances went actively inside buildings to look for Civilians. By doing this, they were quickly loosing health. So it became exciting who would win the final, because the points they won with more efficient firefighters they lost with the daring behavior of their ambulances (differences in score of 0.07). They won the game when finally their hero-ambulances rescued a Civilian that died in our simulation (difference in score of 0.97).

Conclusion

The participation in the German Competition was a real experience. The organization was really good, and allowed the participants to work 24 hours/day. The current simulation environment is not very stable, we needed a rerun for every game.

Recommendations:

1. The performance of the simulator drops exponentially when the fire spreads, causing a lot of problems. Somebody has to look at the fire-simulator.

2. Rule 25 'Valid Game' has to be more specific. There are two possibilities for a rerun:

1. A rerun with the same code, but with different configuration (computing resources, random seed, cycle time).

2. A rerun with the same configuration with modified code (reduced communication, reduced computation).

When you allow the rerun of one team, you must allow the other teams a rerun.

In case 1 this has to be done with the same configuration. In case 2 you have to put a deadline on the debugging (one hour?).

3. Rule 31 'Agents' is too restrictive for other maps than Kobe/VC. Yet, the number of agents allowed is already high. Foligno has 52.776 inhabitants, which is comparable with the Dutch city Capelle a/d IJssel (65.000 inhabitants). The firebrigade of Capelle has 10 vehicles, half of them to be used to battle fire.