Epistemic Status: True story
This post intends to begin the sequence Zbybpu’f Nezl.
In college I once took a class called Rational Choice. Because obviously.
Each week we got the rules for, played and discussed a game. It was awesome.
For the grand finale, and to determine the winner of the prestigious Golden Shark Award for best overall performance, we submitted computer program architectures (we’d tell the professor what our program did, within reason, and he’d code it for us) to play The Darwin Game.
The Darwin Game is a variation slash extension of the iterated prisoner’s dilemma. It works like this:
For the first round, each player gets 100 copies of their program in the pool, and the pool pairs those programs at random. You can and often will play against yourself.
Each pair now plays an iterated Nash bargaining game, as follows. Each turn, each player simultaneously submits an integer number from 0 to 5. If the two numbers add up to 5 or less, each player earns points equal to their own number. If the two numbers add up to 6 or more, neither player gets points. This game then lasts for a large but unknown number of turns, so no one knows when the game is about to end; for us this turned out to be 102 turns.
Each pairing is independent of every other pairing. You do not know what round of the game it is, whether you are facing a copy of yourself, or any history of the game to this point. Your decision algorithm does the same thing each pairing. You do know the results of previous turns in the same pairing.
At the end of the round, all of the points scored by all of your copies are combined. Your percentage of all the points scored by all programs becomes the percentage of the pool your program gets in the next round. So if you score 10% more points, you get 10% more copies next round, and over time successful programs will displace less successful programs. Hence the name, The Darwin Game.
Your goal is to have as many copies in the pool at the end of the 200th round as possible, or failing that, to survive as many rounds as possible with at least one copy.
If both players coordinate to split the pot, they will score 2.5 per round.
To create some common terminology for discussions, ‘attack’ or means to submit 3 (or a higher number) more than half the time against an opponent willing to not do that, and to ‘fold’ or ‘surrender’ is to submit 2 (or a lower number) more than half the time, with ‘full surrender’ being to always submit 2. To ‘cooperate’ is to alternate 2 and 3 such that you each score 2.5 per round.
In this particular example we expected and got about 30 entries, and I was in second place in the points standings, so to win The Golden Shark, I had to beat David by a substantial amount and not lose horribly to the students in third or fourth.
What program do you submit?
(I recommend actually taking some time to think about this before you proceed.)
Some basic considerations I thought about:
1. The late game can come down to very small advantages that compound over time.
2. You need to survive the early game and win the late game. This means you need to succeed in a pool of mostly-not-smart programs, and then win in a pool of smart programs, and then in a pool of smart programs that outlasted other smart programs.
3. Scoring the maximum now regardless of what your opponent scores helps you early, but kills you late. In the late game, not letting your opponent score more points than you is very important, especially once you are down to two or three programs.
4. In the late game, how efficiently you cooperate with yourself is very important.
5. Your reaction to different programs in the mid game will help determine your opponents in the end game. If an opponent that outscores you in a pairing survives into the late game, and co-operates with itself, you lose.
6. It is all right to surrender, even fully surrender, to an opponent if and only if they will be wiped out by other programs before you become too big a portion of the pool, provided you can’t do better.
7. It is much more important to get to a good steady state than to get there quickly, although see point one. Getting people to surrender to you would be big game.
8. Some of the people entering care way more than others. Some programs will be complex and consider many cases and be trying hard to win, others will be very simple and not trying to be optimal.
9. It is hard to tell what others will interpret as cooperation and defection, and it might be easy to accidentally make them think you’re attacking them.
10. There will be some deeply silly programs out there at the start. One cannot assume early programs are doing remotely sensible things.
That leaves out many other considerations, including at least one central one. Next time, I’ll go over my and David’s preparations, and post three will reveal what happened on game night.
Note: Please do not comment here once you have read The Darwin Pregame.