150 Leading with NLP
the computer adversaries a chance to size up the strategy
of the opposing program and counter it if possible. Each
program would collect a history and a reputation, just as
people do. Each program was fed information on how its
adversary had played previous games, so it could change
its strategy based on the opponent’s style (if it was pro-
grammed to do so).
The tournament was marked, the programs scored points
for each round, and the scoring reflected what happens in
life: when both players trust, both do equally well and there
is a win-win situation. But if you trust and they double-cross
you, then you lose and they win. The highest score was for
defection, provided the other program did not also defect
(the equivalent of informing and going free with a reward).
The next highest score was for co-operation, the equivalent
of both players being released from jail for not informing.
Fewer points were awarded if both players defected, as both
would go to jail, but at least neither was fined. Finally the
lowest score was for co-operating when the other player in-
formed. Then you were jailed and had to pay a fine as well.
Fourteen programs were submitted, ranging from simple
strategies like ‘always defect’ or ‘always co-operate’ to ex-
tremely complex ones. One strategy came out a clear winner
at the end of the tournament. Can you guess what it was?
The ‘nice guy’ programs did not win. Indiscriminate co-
operation and trust did not score highly, and I wouldn’t
recommend them in real life either. But neither did the non-
cooperative programs. The winning program was submitted
You defect You co-operate
Other player Both score Opponent scores
defects the same (low) well
You score badly
Other player You score well Both score the
co-operates They score badly same (high)