This game continues from the alternate timeline here where I made two mistakes in the game engine.

  • Bots were passed their own previous move and told it was their opponent's previous move.
  • Bots were always given 0 as the round index instead of the correct positive integer.

CloneBots

Multiple people have noted that CliqueZviBot is outperforming the other CloneBots. This is due to how the CloneBot code interacts with the bugs in the my engine.

The CloneBots still cooperate, but they do so imperfectly. All CloneBot pairings result in 200-300 splits instead of 250-250 splits. The CloneBots use source code parity combined with round number parity to determine who wins the 200-300 split. Therefore if CloneBotA and CloneBotB get a 200-300 split in favor of CloneBotB then they will always get a 200-300 split in favor of CloneBotB.

Rounds 11-30

Round 11

Looking at the obituary I suspect that CooperateBot may not last much longer.

Prediction by Larks after seeing the results from Rounds 1 to 10

Larks' CooperateBot died on round 11.

Round 12

PasswordBot from Team Multics died along with "Why can't we all just get along" from Chaos Army and an NPC.

Round 13

No casualties.

Round 14

6 bots died.

  • BeauBot, OscillatingTwoThreeBot, RandomOrGreedyBot and SimplePatternFinderBot from Chaos Army
  • "Definitely Not Collusion Bot" from Team Multics. Multicore's fodder has been consumed. Team Multics contains only the MimicBot from here on.
  • 1 NPC

Round 15

5 bots died.

  • Silly Invert Bot 2, AttemptAtFair, Insum's CooperateBot, MeasureBot and "Random-start-turn-taking" from Chaos Army
  • 1 NPC

MeasureBot had succeeded in infecting AbstractSpyTreeBot's move method and replacing it with return 0. AbstractSpyTreeBot ought perform better with MeasureBot out of the game.

Round 16

3 bots from Chaos Army died

  • "Silly Counter Invert Bot"
  • LiamGoddard
  • "Pure TFT"

Rounds 17-22

4 NPCs died

Round 23

BendBot and Copoperater [sic] died. BendBot belonged to Zvi. CliqueZviBot does not actually belong to Zvi. It is named after Zvi's strategy from the original name.

Round 24

No casualties.

Round 25

Copybot Deluxe died.

Round 26

RaterBot died. RaterBot performed semantic analysis on its opponents' source code. This may have contributed to breaking the symmetry of the clones.

Round 27

No casualties.

Round 28

Empiricist died. Empiricist was the most complicated bot I agreed to write the code for. The bot is exemplar of a precise, well-written spec of a clever algorithm.

On the first round, Empiricist plays 2. On any subsequent round, it examines the history against the current opponent so far. Let's denote it , where  are Empiricist's plays,  are the opponent's plays and  is the number of rounds played so far. The algorithm is as follows

Step 0: Compute Empiricist's total score so far (denote ) and the opponent's total score so far (denote ). If , then Empiricist plays 3. Otherwise, continue to the following steps.

Step 1: Compute the maximal number  s.t. the last  rounds of the game are a repetition of some previous sequence. That is,  is maximal s.t. there exists  with  s.t. the sequence  is identical to the sequence . If no  satisfies this property, set .

Step 2: Find the latest subsequence among previous repetitions, that is, the maximal  that satisfies the property above w.r.t. the chosen m. If , set .

Step 4: Examine . If , Empiricist plays 5-y. If , Empiricist plays 2.

Round 29

1 NPC died.

Round 30

No casualties.

Summary of Rounds 11-30

Everything so far

List of Survivors

The CloneBots and the MimicBot are all still alive.

BotPopulation
EarlyBirdMimicBot1013
Akrasia Bot950
A Very Social Bot916
CliqueZviBot907
Clone wars, episode return 3888
a_comatose_squirrel423
incomprehensibot221
KarmaBot48
CloneBot8

Two bots from Chaos Army survived this first ¼ of Order 66.

BotPopulation
AbstractSpyTreeBot15
Winner against low constant bots5

One NPC survived too. Silly 2 Bot always returns 2.

BotPopulation
Silly 2 Bot5

In early tests of the game, I discovered that sometimes bots got stuck at a population of 2 from which they never died nor recovered. I added custom code to finish off any bot with a population of 2 or less. Any bot whose population drops to 2 or less will die.

Multicore's Mystery

Multicore is in first place. But Multicore should not just be in first place. As the sole traitor among the CloneBots, Multicore's MimicBot should be dominating this competition. Plus, as a simulator, it should be able to cooperate on the first turn despite receiving misinformation about its opponent's previous move. I suspect that the MimicBot's simulator is useful because the other simulator, AbstractSpyTreeBot has the highest population of all non-clones.

Multicore's failure to completely dominate probably has something to do with the bugs in the game engine. But there's something else which at play too.

Simple bots are most advantageous to simulate. They are easy to maximize cooperation with and there is little danger of simple bots winning in the long game. The simplest bots were my silly bots. The next simplest bots were the bots I wrote on behalf of non-programmers. I programmed exclusively in Lisp. The MimicBot only has has code to simulate opponents written in Python3 and cannot simulate bots written in Lisp. (The same goes for AbstractSpyTreeBot, which MimicBot's simulator came from.) Therefore MimicBot cannot simulate the bots which it would be most worthwhile to simulate.

If this is true then AbstractSpyTreeBot continues to influence this game.

Today's Obituary

BotTeamSummaryRound
CooperateBot [Larks]Chaos Army"For the first 10 turns: return 3. For all subsequent turns: return the greater of 3 and (5 - the maximum value they have ever submitted)"11
PasswordBotMulticsFodder for EarlyBirdMimicBot12
Why can't we all just get alongChaos ArmyDoesn't negotiate with terrorists. Doesn't overly punish slackers. Attempts to establish steady tit-for-tat.12
Silly TFT Bot 3NPCsTit-for-Tat starting at 312
Silly Cement Bot 2-3NPCsReturns 2 or 3 on the first turn. Otherwise, returns 5 - opponent_first_move.14
BeauBotChaos ArmyAt 528 lines, this is the most sophisticated bot to die so far. It picks one of 3 simple strategies based on it's opponent's behavior. It also adjusts its behavior based on the round.14
OscillatingTwoThreeBotChaos Army"cooperates in the dumbest possible way"14
Definitely Not Collusion BotMulticsColludes with EarlyBirdMimicBot14
RandomOrGreedyBotChaos ArmyIf the opponent averaged less than 2.5 over the last 100 turns then plays int(5 - opponent_avg). Otherwise randomly selects 3 or 2 randomly.14
SimplePatternFinderBotChaos ArmyFinds simple patterns.14
Silly Invert Bot 2NPCsStarts with 2. Then always returns 5 - opponent_previous_move15
AttemptAtFairChaos ArmyOscillates between 3 and 2, starting with 3.15
CooperateBot [Insub]Chaos ArmyLet MLM = my last move, OLM = opponent's last move. On the first turn, play 2. On subsequent turns: [Fork 1] If (MLM + OLM = 5), then play OLM [Fork 2] Otherwise, flip a coin and play max(MLM, OLM) with 50% probability, and (5 - max(MLM, OLM)) with 50% probability15
MeasureBotChaos armyAttempts to hijack a simulator's move method and return 0. This succeeded against AbstractSpyTreeBot and failed on EarlyBirdMimicBot. Otherwise, it uses a hand-coded decision tree with 20 terminal leaves.15
Random-start-turn-takingChaos ArmySelects 3 or 2 randomly until symmetry is broken. Then oscillates between 2 and 3.15
Silly Counter Invert BotChaos ArmyStarts by randomly playing 2 or 3. Then always returns 5 -opponent_previous_move.16
LiamGoddardChaos ArmyStarts with 3 2 3 2. Then picks one of 5 strategies to use for the rest of the game.16
Pure TFTChaos Army"For the first round, play 2 or 3 with a 50/50 chance of each. For each subsequent round, play whatever the opponent played on the previous round."16
Silly Random Invert Bot 2-3NPCsStarts by randomly playing 2 or 3. Then always returns 5 -opponent_previous_move. (Same as Silly Counter Invert Bot.)17
Silly Invert Bot 3NPCsStarts with 3. Then always returns 5 - opponent_previous_move19
Silly Cement Bot 3NPCsReturns 3 on the first turn. Otherwise, returns 5 - opponent_first_move.20
Silly TFT Bot 2NPCsTit-for-tat, starting at 2.21
BendBotChaos ArmyFirst proposal was rejected as too complicated. Second proposal was rejected as too complicated. Third proposal was accepted. For details, see Zvi's write-up here.23
Copoperater [sic]Chaos ArmyTit-for-tat, starting at 2.23
CopyBot DeluxeChaos ArmyTit-for-tat. Picks starting value of 2 or 3 based off of round number.25
RaterBotChaos ArmyEstimates opponent's aggression by counting the number of 3s, 2s, return 3s and return 2 instances in its source code. Then picks a strategy based off of that.26
EmpiricistChaos ArmyPerforms the best strategy that would have worked against historical data.28
Silly Cement Bot 3NPCsReturns 2 on the first turn. Otherwise, returns 5 - opponent_first_move.29

The mutant game will continue on November 27, 2020.

New Comment
2 comments, sorted by Click to highlight new comments since:
  • Bots were passed their own previous move and told it was their opponent's previous move.
  • Bots were always given 0 as the round index instead of the correct positive integer.

I was wondering what the second mistake was - I notice that I am no longer confused :)

Originally I was confused in this game as to why CliqueZviBot started decreasing around round 23 but I guess it is a combination of:

  • getting 300-200 against a_comatose_squirrel, incomprehensibot and KarmaBot which have become a smaller part of the pool.
  • getting 200-300 against A Very Special Bot, Clone wars, episode return 3 and EarlyBirdMimicBot which are becoming a larger part of the pool

(not all of them but probably most, especially the first bullet)

In the Mutant Game, the PasswordBots aren't really on my team. Since it appears to them that the opponent started with 3, they will play 3 on every turn against everybody.

If the round index is always 0, that means the clone truce never ends and you have population dynamics based on which clones are getting 300-200 against which other clones. Not sure if it will be stable or not.

I suspected that my simulation strategy didn't end up being all that useful, but I'm still curious what bots I managed to simulate at all. Presumably at least the PasswordBots. I guess I can find out when the code is released.