To LUGNET HomepageTo LUGNET News HomepageTo LUGNET Guide Homepage
 Help on Searching
 
Post new message to lugnet.roboticsOpen lugnet.robotics in your NNTP NewsreaderTo LUGNET News Traffic PageSign In (Members)
 Robotics / 24701
24700  |  24702
Subject: 
Re: Design
Newsgroups: 
lugnet.robotics
Date: 
Thu, 8 Dec 2005 21:40:34 GMT
Original-From: 
steve <sjbaker1@airmail.SAYNOTOSPAMnet>
Viewed: 
1198 times
  
danny staple wrote:
I do remember seeign something very similar to this during my studies,
except it was a genetic alogorithm based on rounds in an arena. The
genomes would gain fitness by being able to push their opponent out of
the ring. Some of the most sucessful were just vast blocks which would
just fall over.

Another story like that - relating to neural networks - is of a busy
Japanese subway system where they wanted to prevent overcrowding of
platforms at the stations.

The story goes that they planned to hook up a camera with a computer
observing the station using a neural network algorithm in order to
decide when to shut the turnstiles and prevent more people entering
the platform.

Training the system on photographs of full and empty platforms allowed
the neural network to 'learn' a solution to their problem without anyone
having to write tedious image processing software.

The system went into operation and worked perfectly for many months
until on a public holiday - it unaccountably locked the turnstiles
when there was hardly anyone standing on the platform.

After much careful analysis of the resulting neural net, they discovered
that rather than triggering when there were too many people present,
the neural net had learned how to read the clock on the wall opposite
the camera - and was in fact shutting the doors at the same rush-hour
times every day - regardless of the number of people on the platform!

I don't recall where I heard this story - so it's very likely that I
have some of the details wrong.

However, it nicely illustrates the problems of learning/evolving
systems.



Message has 1 Reply:
  Those Silly AI programs (was: Design)
 
My favorite: A neural net was trained to recognize tanks. It worked perfectly on all the test data, but failed completely on real data. To create the test data, they shot pictures of some tanks, and some pictures of cars and other non-tank vehicles (...) (19 years ago, 9-Dec-05, to lugnet.robotics)

Message is in Reply To:
  Re: Design
 
I do remember seeign something very similar to this during my studies, except it was a genetic alogorithm based on rounds in an arena. The genomes would gain fitness by being able to push their opponent out of the ring. Some of the most sucessful (...) (19 years ago, 8-Dec-05, to lugnet.robotics)

6 Messages in This Thread:

Entire Thread on One Page:
Nested:  All | Brief | Compact | Dots
Linear:  All | Brief | Compact

This Message and its Replies on One Page:
Nested:  All | Brief | Compact | Dots
Linear:  All | Brief | Compact
    

Custom Search

©2005 LUGNET. All rights reserved. - hosted by steinbruch.info GbR