the unreasonable effectiveness of copying

classic Classic list List threaded Threaded
2 messages Options
Reply | Threaded
Open this post in threaded view
|

the unreasonable effectiveness of copying

Roger Critchlow-2
Last month Science reported the results (http://www.sciencemag.org/cgi/content/abstract/sci;328/5975/208) of a social learning tournament (http://www.intercult.su.se/cultaptation/tournament.php).  

In non-social algorithmic learning, an agent has the choice between exploiting what it already knows and exploring to discover new behaviors.  In social learning there is a third choice, to observe your peers and learn from their behavior.

The tournament, implemented in matlab, evolved populations of social learning algorithms according to their success in exploiting a 100-arm bandit whose payoffs changed over time.  Each learner chose to exploit a previously learned result, explore a new result, or observe a peer on each turn.  Learners died after an average of 50 turns and were replaced by new learners according to the success of the algorithms.  New learners inherited their species algorithm but no experience.  Personal exploration yielded exact results, but the results of social observation was blurred by observational error.  The tournament ran a round robin of head to head competitions followed by a series of melee rounds among the ten best strategies.

What happened surprised everyone.  The best strategies were biased more than 95% toward observation over exploration.  The differences between the top strategies were largely in the triggers to switch away from exploitation in response to changes in the payoffs.

So, as I compose this message I am tempted to include the Science article that reports the results, but that would violate the copyright of the AAAS.  If you are inside the AAAS pay wall, then you may observe what this tournament learned, otherwise you will remain ignorant until you cough up the fee or find a public library or find someone else who will break the wall.

Our intellectual property laws are formulated as brakes on social learning.  They restrict the ability of people to learn from what is known and to use what they learn.  As formulated they prevent us, as a species, from using what we, as a species, have learned.   Perhaps it's time to reverse the formulation.  Rather than granting monopolies to "inventors" and "creators" grant unrestricted licenses to learn and to copy.  Rather than negotiating licences and rights on a case by case basis, let every manufacturer and content publisher pay a "prior art" tax and let the lawyers fight over the appropriateness of the rate and the distribution of the proceeds amongst the prior artists.

-- rec --


============================================================
FRIAM Applied Complexity Group listserv
Meets Fridays 9a-11:30 at cafe at St. John's College
lectures, archives, unsubscribe, maps at http://www.friam.org
Reply | Threaded
Open this post in threaded view
|

Re: the unreasonable effectiveness of copying

Victoria Hughes
very interesting. thanks.
Tory


On May 3, 2010, at 12:29 PM, Roger Critchlow wrote:

Last month Science reported the results (http://www.sciencemag.org/cgi/content/abstract/sci;328/5975/208) of a social learning tournament (http://www.intercult.su.se/cultaptation/tournament.php).  

In non-social algorithmic learning, an agent has the choice between exploiting what it already knows and exploring to discover new behaviors.  In social learning there is a third choice, to observe your peers and learn from their behavior.

The tournament, implemented in matlab, evolved populations of social learning algorithms according to their success in exploiting a 100-arm bandit whose payoffs changed over time.  Each learner chose to exploit a previously learned result, explore a new result, or observe a peer on each turn.  Learners died after an average of 50 turns and were replaced by new learners according to the success of the algorithms.  New learners inherited their species algorithm but no experience.  Personal exploration yielded exact results, but the results of social observation was blurred by observational error.  The tournament ran a round robin of head to head competitions followed by a series of melee rounds among the ten best strategies.

What happened surprised everyone.  The best strategies were biased more than 95% toward observation over exploration.  The differences between the top strategies were largely in the triggers to switch away from exploitation in response to changes in the payoffs.

So, as I compose this message I am tempted to include the Science article that reports the results, but that would violate the copyright of the AAAS.  If you are inside the AAAS pay wall, then you may observe what this tournament learned, otherwise you will remain ignorant until you cough up the fee or find a public library or find someone else who will break the wall.

Our intellectual property laws are formulated as brakes on social learning.  They restrict the ability of people to learn from what is known and to use what they learn.  As formulated they prevent us, as a species, from using what we, as a species, have learned.   Perhaps it's time to reverse the formulation.  Rather than granting monopolies to "inventors" and "creators" grant unrestricted licenses to learn and to copy.  Rather than negotiating licences and rights on a case by case basis, let every manufacturer and content publisher pay a "prior art" tax and let the lawyers fight over the appropriateness of the rate and the distribution of the proceeds amongst the prior artists.

-- rec --

============================================================
FRIAM Applied Complexity Group listserv
Meets Fridays 9a-11:30 at cafe at St. John's College
lectures, archives, unsubscribe, maps at http://www.friam.org

-----------------------------------


============================================================
FRIAM Applied Complexity Group listserv
Meets Fridays 9a-11:30 at cafe at St. John's College
lectures, archives, unsubscribe, maps at http://www.friam.org