ICFP 2008 post-mortem

This year's ICFP contest was a traditional one: Write some code that solves an optimization problem with finite resources, debug it using sample data sets, send it in, and the judging team will run it on secret (presumably more difficult) data sets, and see whose program does the best. The problem was to create a control program for an idealized Martian rover that had to drive to home base while avoiding craters, boulders, and moving enemies.

I read the problem description at noon on Friday, but didn't have time to work on the contest until Saturday morning.

The first task was to choose a language. On the one hand, the strict time limit argued for an easy-to-hack "batteries included" language like Python, for which libraries, IDEs, and cross-platform runtime were all readily available. On the other hand, the requirement for high performance and ability to correctly handle unknown inputs argued for a type safe, compiled language like ML or O'Caml.

I spent a half an hour trying to set up an O'Caml IDE under Eclipse, but unfortunately was not able to figure out how to get the debuger to work. Then I switched to Python and the PyDev IDE, and never ran into a problem that made me consider switching back.

I realize that the resulting program is much slower than a compiled O'Caml would be, and it probably has lurking bugs that the O'Caml type system would have found at compile time. But it's the best I could do in the limited time available for the contest.

It was very pleasant to develop in Python. It's got a very nice syntax. I was never at a loss for how to proceed. Either it "just worked", or else a quick web search would immediately find a good answer. (Thanks Google!)

The main drawback was that the Python compiler doesn't catch simple mistakes like uninitialized variables until run time. Fortunately that wasn't too much of a problem for this contest, as the compile-edit-debug cycle was only a few seconds long, and it only took a few minutes to run a whole test suite.

The initial development went smoothly: I wrote was the code to connect to the simulation server and read simulation data from the server. Then I created classes for the various types of objects in the world, plus a class to model the world as a whole. I then wrote a method that examined the current state of the world and decided what the Martian rover should do next. Finally I wrote a method that compared the current and desired Martian rover control state, and sent commands back to the simulation server to update the Martian rover control state.

The meat of the problem is deciding how to move the rover. The iterative development cycle helped a lot here -- by being able to run early tests, I quickly discovered that the presence of fast-moving enemies put a premium on high speed movement. You couldn't cautiously analyze the world and proceed safely, you had to drive for the goal as quickly as possible.

My initial approach was to search for the closest object in the path of the rover, and steer around it. This worked, but had issues in complicated environments. Then I switched to an idea from Craig Reynolds' Not Bumping Into Things paper: I rendered the known world into a 1D frame buffer, and examined the buffer to decide which way to go. That worked well enough that I used it in my submission.

I spent about fourteen hours on the contest: Two hours reading the problem and getting the IDE together, ten hours over two days programming and debugging, and about two hours testing the program on the Knoppix environment and figuring out how to package and submit the results.

Things I wish I had had time to do
  • My rover is tuned for the sample data sets. The organizers promised to use significantly different data sets in the real competition. Unfortunately, I didn't have time to adapt the program to these other data sets, beyond some trivial adjustments based on potential differences in top speed or sensor range.
  • I model the world at discreet times, and don't account for the paths objects take over time. I can get away with this because I'm typically traveling directly towards or away from important obstacles, so their relative motion is low. But I would have trouble navigating through whirling rings of Martians.
  • I don't take any advantage of knowledge of the world outside the current set of sensor data. The game explicitly allows you to remember the world state from run to run during a trial. This could be a big win for path planning when approaching the goal during the second or later trials.
  • I don't do any sort of global path planning. A simple maze around the goal would completely flummox my rover.
I very much enjoyed the contest this year. I look forward to finding out how well I did, as well as reading the winning programs. The contest results will be announced at the actual ICFP conference in late September.

Comments

Popular posts from this blog

GLES Quake - a port of Quake to the Android platform

A Taipei-Torrent postmortem: Writing a BitTorrent client in Go

A Multi-threaded Go Raytracer