GlyphChess War Stories

• A major reason for GlyphChess development was that it was really quiet at work over the holidays and I had a broken leg so I couldn't go skiing.
• The first version of GlyphChess was developed within a single uninterrupted day (and night). It consisted of a few crude shell scripts and was really hacked together. Board scans were somewhat unreliable and took several seconds, as it was doing thousands of program executions per scan.
• The C program that replaced those scripts is about 600 lines of code and can execute before the scan bar has had time to return to the home position. It consists mainly of a set of calls to a math library and the DataGlyph toolkit, plus the chess piece logic. Interaction with the chess engine is still scripted. Board readout is quite reliable.
• I noticed that the linear programming library lp_solve was slightly out of date and also slightly mispackaged in the Debian Linux distribution. I filed two bugs on Jan 6, 2003, and the maintainer had them both fixed by Jan 12, 2003. Shortly thereafter, I dumped lp_solve for a linear least squares routine in Numerical Python.
• Several changes were made to the DataGlyph Toolkit. As expected, now every detected glyph mark reports both its logical and pixel position. An unexpected change was modifying the address carpet decode algorithm. Address voting now requires a quorum, which helps when you have a circular capture window such as the base of a chess piece. Several minor bugs were encountered in the toolkit while developing GlyphChess. They were filed and then fixed or worked around.
• I originally marked the pieces with GlyphTones matching the piece itself. That went away as I kept tweaking things to improve reliability, as our cell size and angular detection modules were being stressed out. Switching to the serpentone detectors produced better results, so GlyphTones were re-introduced.
• I was hoping for incredible extrapolated registration results, perhaps down to pixel level accuracy. That didn't happen, although there was more than enough accuracy for GlyphChess. Further investigation revealed accuracy scales dramatically with scan resolution. Scans at 400 dpi have significantly less registration error than 200 dpi scans.
• The project plan consisted of a simple diagram of possible improvements, and was drawn up the day after the initial hacked version of GlyphChess was written. Tasks were tackled on an as needed basis until I was satisfied that the system worked well enough for demonstration. I have since lost the sheet of paper. Eventually that process degraded into simply doing a ton of scans and then analyzing any problems that were encountered.
• I lost the black king, so I substituted a random object that was lying around with the appropriate DataGlyph attached to the bottom. Works like a charm.
• Thank goodness there wasn't a robot arm laying around nearby, or I would probably still be working on the robot arm variant, complete with knocking the pieces on the floor as a sore loser.
• GlyphChess is a surprisingly fun demo. I really enjoy physically setting up a position on the board and letting the computer chess engine duke it out on screen. I wish my desktop scanner was a little bigger.
• The crafty chess engine has a machine learning capability. As described in the manual, "Currently this is a 3-bit boolean switch, bit 1 (001) controls book learning, bit 2 (010) controls position learning, and bit 3 (100) controls result learning." With the 1983 movie WarGames in mind, I ran GlyphChess overnight to see if it would evolve into a superhuman AI intent on conquering the world. As I cautiously entered my office the next day, I discovered that the computer had spent the entire time learning more about chess and was as fixated on the game as ever.