lack of separation in classifier systems doug wampler thursday, 09 november 2000

13
Lack of Separation in Classifier Systems Doug Wampler Thursday, 09 November 2000

Upload: jodie-bradford

Post on 19-Jan-2016

213 views

Category:

Documents


0 download

TRANSCRIPT

Page 1: Lack of Separation in Classifier Systems Doug Wampler Thursday, 09 November 2000

Lack of Separation in Classifier Systems

Doug Wampler

Thursday, 09 November 2000

Page 2: Lack of Separation in Classifier Systems Doug Wampler Thursday, 09 November 2000

What is a Classifier System (CFS)?

• A Classifier System is a cognitive system that is able to classify or identify events within it’s environment and then react to these events.

• In order to construct a classifier system, we need (1) the environment, (2) rules to classify/identify events, (3) actions to react to the events, and (4) the system itself.

Page 3: Lack of Separation in Classifier Systems Doug Wampler Thursday, 09 November 2000

General Classifier System Problems

• Classifier Systems have three major problems (1) deception, (2) mutlimodality, and (3) lack of separation. (1) & (2) are strongly related and combine to form generalized deception.

• We are selecting the simpler problem, lack of separation.

Page 4: Lack of Separation in Classifier Systems Doug Wampler Thursday, 09 November 2000

Problem Description

• For a Classifier System, below some critical distance, two different rules are considered to be similar. If both of these rules are necessary to solve a particular problem, the Classifier System will be unable to stably maintain both rules.

• In our study, we will call this phenomenon Si, which represents the minimum number of similar rules necessary to solve a particular problem.

Page 5: Lack of Separation in Classifier Systems Doug Wampler Thursday, 09 November 2000

Research Objective

• Predict & prevent lack of separation by devising a method for obtaining Si for any given Classifier System, where Si represents the minimum number of similar rules necessary to solve a particular problem.

• For example, our future Mars probe has two similar rules, “Approach Velocity” and “Orbital Velocity”.

Page 6: Lack of Separation in Classifier Systems Doug Wampler Thursday, 09 November 2000

Research Objective

• If our probe removes “orbital velocity” because it is believed to be redundant (based on predicate comparison), and substitutes “approach velocity” instead, we have lack of separation and failure.

• Our research will attempt to find an accurate method of finding the value of Si.

Page 7: Lack of Separation in Classifier Systems Doug Wampler Thursday, 09 November 2000

Literature Review

• Comp.ai.genetic FAQ by Jorg Heitkotter & David Beasley. Definition of CFS, other resources.

• What Makes a Problem Hard for a Classifier System? David E. Goldberg, Jeffrey Horn, Kalyanmoy Deb. Major problems in CFS design.

Page 8: Lack of Separation in Classifier Systems Doug Wampler Thursday, 09 November 2000

Importance of Research

• If lack of separation could be conquered, Classifier Systems could :

• Detect & repair the faults that keep causing our Mars probes to disappear;

• Cause bad software to become perfect software over a long period of time, with no human intervention;

• Cause machine learning to happen exponentially faster than human learning.

Page 9: Lack of Separation in Classifier Systems Doug Wampler Thursday, 09 November 2000

Research Design

• Contrive several situations in which lack of separation occurs;

• Calculate the specific point in each case where lack of separation occurs;

• Derive the specific value of Si for each instance, where Si = the minimum number of similar rules to solve a particular problem.

• Goal – predict & prevent lack of separation.

Page 10: Lack of Separation in Classifier Systems Doug Wampler Thursday, 09 November 2000

Schedule

• March 2001 - Devise specific occurrences of lack of separation problems

• June 2001 - Develop methods for avoiding lack of separation in the above examples

• September 2001 - Isolate value of Si in each avoidance example above.

• November 2001 – Derive general formula for Si.

• December 2001 - Publish research abstract.

Page 11: Lack of Separation in Classifier Systems Doug Wampler Thursday, 09 November 2000

Facilities & Special Resources

• Conservative Estimate - No facilities or special resources necessary.

• Aggressive Estimate – UNIX server and 2-3 UNIX workstations necessary for design & implementation of CFS software prototype.

Page 12: Lack of Separation in Classifier Systems Doug Wampler Thursday, 09 November 2000

Deliverables

• Research abstract containing detailed examples of occurrences of lack of separation problems, accompanied by examples of problem avoidance using Si variable isolated by our research, -or-

• Research abstract above, plus CFS software product containing these examples as well.

Page 13: Lack of Separation in Classifier Systems Doug Wampler Thursday, 09 November 2000

Conclusion

• Questions?

• The End