Talk:The Evolution of Cooperation

From Wikipedia, the free encyclopedia

	This article is part of WikiProject Game theory, an attempt to improve, grow, and standardize Wikipedia's articles related to Game theory. We need your help! Join in \| Fix a red link \| Add content \| Weigh in

Start	This article has been rated as start-Class on the assessment scale.
Low	This article is on a subject of low-importance within game theory.

I don't know about U.S. publications but the UK penguin edition is entitled 'The Evolution of Co-operation' as opposed to 'The Evolution of Cooperation' as it is in this article. MagicBez 17:49, 12 August 2005 (UTC)

[edit] Criticism: Neglects recognition/assessment of value of information

This is one of my favourite books with all sorts of potential for pragmatic ethics. One thing troubles me though: the exposition never addressed the question of the extent of knowledge available to the algorithms. Tit-for-Tat manages with the minimum on information; namely just he record of the last encounter with its (recognised) opponent. Its certainly possible that better results might be achieved by an algorithm with access to more complete information, such as:

1) record of encounters with this opponent
2) record of opponent's encounters (including with other participants)
3) record of encounters with all opponents
4) record of all encounters between any combination of opponents
5) points already acquired by the opponent
6) opponent's source code.

6 is an extreme case where it is very obvious that the information would lead to improved results. The value of information in other case, while certainly less blatant, can still be supported by compelling arguments:

2 may be partially comparable to 6, in that in many cases it may be possible to reverse engineer a opponent's algorithm. Also, some combinations of the information above would add to an algorithms's knowledge of the make up of the algorithm population. For instance, it would help in quickly establishing a reliable assessment of the proportion of nice algorithms (cf Axelrod's use of this term), allowing one's own algorithm to make global adjustments as appropriate. 5 is useful as an input to decisions on disruptive behaviour ("stop the leader" - in games where a algorithm is more interested in finishing first than in attaining a high point total) and might also serve as a single parameter estimate of an opponent's non-cooperativeness and/or exploitability.

The potential for using information about the opponent algorithm's performance in a variety of ways to enhance one's own algorithm's performance seems fairly transparent, and it is surprising that this is neglected. Another interesting field of study would be to assess the value of certain types of information. What would be a fair handicap for an algorithm to accept (in negative starting points) for certain types of information delivered at certain times.

The above comments, intended as criticisms on Axelrod's book, may also qualify as original research. Hence my reluctance to introduce them into the article. However I'd very much welcome responses from Wikipedians with similar interests/insights..

--Philopedia 18:00, 4 January 2007 (UTC)