Real classromm about SD
Artificial classroom  about MAXCSP (inhabited by agents)
common:
statements, get questioned, defended and discounted,
reputations go up and down.

Mini-Classroom metaphor

SCG creates a mini-classroom for the agents to learn about topic X.
The agents make statements about X which have the flavor
of challenges: some of those statements may be discountable.
The agents offer and accept challenges, each having a confidence in [0,1].
Each agent starts with an initial reputation. An agent
wins reputation (proportional to challenge confidence and
offerer reputation) when it successfully discounts a challenge
of another agent. Or an agent wins reputation (proportional to
challenge confidence and acceptor reputation) when it
successfully supports its own challenge after it was accepted
by another agent.
The game is a zero-sum game with respect to reputation.

Discounting and supporting of a challenge involves a constructive,
interactive protocol that produces objects which are given to a  predicate,
either resulting in discounting or supporting.

The agent with the highest reputation wins the game.
When X is algorithmic, the game is about improving algorithms.
When X is "optimization algorithms in domain Y
whose decision version is NP-complete (NPO for Y)",
e.g. MAXCSP, the winning agent has the best algorithms for MAXCSP
and the challenge language that was chosen.
Different challenge languages bring about different algorithms.

SCG(X) creates both a competition and and a collaboration between the agents.
The competition is about reputation and the collaboration is about
the constructive feedback the agents give to each other.
When an agent loses reputation, it gains information that points to
a bug in its algorithm or a conceptual problem based on a lack of knowledge.

The SCG games run over many rounds and produce a wealth of data
from which useful information needs to be extracted.
Currently the agents are improved by humans but there is an
opportunity to use machine learning.

We have SCG implement on the web, using a client-server architecture.
It is used with MAXCSP as actual parameter in my software development classes.

http://www.ccs.neu.edu/home/lieber/evergreen/specker/scg-home.html


=================================
The SCG is similar to what happens in a class room
in which subject area X is taught.
With SCG we create a mini class room for the artificial agents
to learn.

The professor expresses statements about X to the students 
which they interpret as challenges (they might be able
to successfully argue against those statements):

The professor has a high reputation and a high confidence
in its statements.
When a student questions a statement and successfully discounts it,
the reputation of the student goes up 
and the reputation of the professor drops.
When a student questions a statement and the professor 
successfully supports it,
the reputation of the professor goes up.

The professor encourages the students to make statements
about subject area X.
Student Silvia, which is believed to be very smart and enjoys
a high reputation, makes a statement with high confidence.
Student Anton questions Silvia's
statement and after some arguing
Anton agrees that Silvia successfully supported her statement.
This lifts Silvia's reputation even higher and the 
reputation of Anton drops.

In SCG, the statements to be made are normed by a challenge language.
The kind of arguing to be done is defined by an interactive
protocol. And whether there is supporting or discounting
is clearly decided by the belief predicate.

In both the class room and the SCG, the goal is to have 
correct statements that cannot be discounted.
In both the class room and the SCG, the goal of the participants is 
to lift their reputation.
In the class room case, the professor will remember 
who participated and who corrected him or another student.

When does the mini class room approach work?
It must be possible to express the knowledge precisely in a program
that can then act out the knowledge. Depends on the power of
the modeling/programming language used.
Challenges must have an uncertainty to them.
Are niches the only way?
Is (Exists/ForAll)+ or (ForAll/Exists)+ the only way?
(Exists problem in niche/ForAll solutions)+
(Exists problem in niche/ForAll algorithms)+
(Exists algorithm/ForAll problems in niche)+


======================================================
http://www.pragprog.com/the-pragmatic-programmer/extracts/tips
Search keywords to be discussed in class:
test
saboteur (how much better is an optimally biased coin over a fair coin?)
coupl -> Law of Demeter
abstraction
single
learning
perfect
exception
critical