The problem with 'random' bots

This article specifically focuses on Vienna bot. Before I begin, I’d like to give some kudos for getting all of the bots operational again.

I’ve done a review of the (RNA) series of puzzles I designed a few weeks back. The RNA (Random Number Association) series of puzzles were created by going to a random number generator and requesting between 100-400 random integers between the values of 1 and 4. The resulting random series would then be translated into sequence data- 1’s were Adenines, 2’s were Uracils, 3’s were Guanines, and 4’s were Cytosines. I would then create an RNA strand of equivalent length, enter the random sequence, and then copy the resulting random structure to create the puzzle. I used these puzzles to ask a few questions, but the big one regarded Vienna. Can random solve random?

Well, the results so far seem to indicate that it can, at random. Out of 14 published and tested so far, Vienna has failed 9 of them. I anticipate that the value of 9 failed is anomalous, and would be 7 had I not published specific subsets of the RNA series that are a little more difficult for Vienna. Truly, I expect, with future tests, that ‘fail’ value will converge on 50 per cent. The reason for this is that both the bot and the puzzle are random, and since the bot may only achieve the values of failure or success, it will fail 50% of the time and succeed 50% of the time against a random RNA.

A little more interesting, though, is Vienna’s performance against the (RNA) - free series, which forbade the random number generator to generate a given base. Vienna failed three out of four of those puzzles, only succeeding on Adenine free. I expect that part of Vienna’s shortcomings may be that it randomizes all bases instead of going about bases meant to be unpaired with more caution (this is likely why SSD bot’s performance on these puzzles is a little better).

Also, thus far, Vienna’s performance against the (RNA) repetition structures seems to be quite poor indeed. Perhaps more on that later.

I’m not trying to knock eteRNA, CMU, or Vienna, but I figured that I’d point out what I believe I’ve observed. After all, the player puzzle designer does serve the purpose of pointing out errors in the algorithms. If I’m wrong about something, please correct me. If there’s anything else anyone else has observed, I’d love to hear it.

I kind of got on a roll and forgot where I was going with that. >_>

Anyway, I’ve also noticed that the bots don’t fare as well against loop-heavy RNAs in general. So there’s that too.