Date post: | 17-Jan-2018 |
Category: |
Documents |
Upload: | cynthia-armstrong |
View: | 222 times |
Download: | 0 times |
Persistent Playgrounds
Fall 2011Managing Software Development
05/03/23 1Persistent Playgrounds
Alternative Introduction
• From an Independence-Friendly Logic website.
05/03/23 Persistent Playgrounds 2
Quantifier Game of Lovers• http://mtl.uta.fi/logic-group/iflogic.html• Bob and Alice are two lovers who can never seem to
agree about anything. Whenever Alice makes a claim, Bob immediately tries to refute it. If Alice makes an existential claim, Bob demands that she identify an individual satisfying it. When Alice asserts a negated formula the lovers switch roles: Bob attempts to defend the claim, whilst Alice tries to refute it.
• When Alice asserts an ordinary first-order claim, the ensuing dispute is a game of perfect information. If the claim is true, Alice will always win. If it is false, Bob will always win.
05/03/23 3Persistent Playgrounds
Proofs / Winning Strategies
• Skolemization is the process of eliminating the existential quantifiers from a formula by introducing fresh function symbols.
• A Skolem function produces a witness for an existential claim. Taken together, a set of Skolem functions for a claim can be seen as encoding a strategy for Alice in the associated semantic game. If the claim is true, the Skolem functions can be assumed to encode a winning strategy.
05/03/23 4Persistent Playgrounds
Persistent Playgrounds
Persistent Playground• Tracks the evolution of
solution techniques in a domain over several years.
• Compares old approaches with new ones. The new approaches should succeed over all old ones.
• Builds incrementally a knowledge base through a sequence of tournaments. Develops the techniques to defend the claims in knowledge base.
Isolated Playground• Persistent playground with
only one tournament.
05/03/23 5Persistent Playgrounds
Comparison
Persistent Playground• tournament• second opinion evaluation• data mining
– strongly defended
Isolated Playground• tournament• nothing• data mining
– two agree
05/03/23 6Persistent Playgrounds
Second Opinion• After a tournament is complete we have:– welfare set from previous tournament– strongly defended claims -> might be true– strongly refuted claims -> might be false
• Apply data mining step. • Many of the strongly defended or refuted claims
have not been seen by many avatars.• All avatars must give their decision (refute,
strengthen or agree) on each strongly defended or refuted claim as well as on the old welfare claims.
05/03/23 7Persistent Playgrounds
Second Opinion
• The second opinion evaluation will change the set of strongly defended and refuted claims.
• The second opinion evaluation is like a tournament but without proposing claims. Instead claims come from data mining.
• Apply data mining step again.
05/03/23 8Persistent Playgrounds
SCG
• Tournament– list of claims with statistics
• Datamining Results of Tournament– WelfareClaims• strongly defended claims
– DemocraticCleanup• claims are even more strongly defended or they are
less strongly defended and dropped from the welfare set. All vote with justification.
05/03/23 9Persistent Playgrounds
Statistics collected for claim
• list of proposers• for each claim and proposer– how often defended• reputation of defender at time of defense
– how often refuted• reputation of refuters at time of refutation
05/03/23 10Persistent Playgrounds
Statistics collected for claim
• for each claim– how often defended• average reputation of defender at time of defense
– how often refuted• average reputation of refuters at time of refutation
05/03/23 11Persistent Playgrounds
Data mining: welfare claimssimple
• C : d: defended r: refuted• d: how often defended• r: how often refuted• Keep it simple– d/(d+r) > 0.99: true welfare claims– r/(d+r) > 0.99: false welfare claims
05/03/23 Persistent Playgrounds 12
Data mining: welfare claimsmore complex
• C : d: defended (ard) r: refuted (arr)• d: how often defended• r: how often refuted• More complex:– d*ard/(d*ard+r*arr) > 0.99: true welfare claims– r*arr/(d*ard+r*arr) > 0.99: false welfare claims
• ard: average reputation of defender• arr: average reputation of refuter
05/03/23 Persistent Playgrounds 13
Why reputation
• if r1>r2: – claim C is defended against an opposer with
reputation r1 counts more than – claim C is defended against an opposer with
reputation r2
05/03/23 14Persistent Playgrounds
Events
• refute– defender• current reputation
– opposer• current reputation
– claim– outcome: refuted/defended
05/03/23 15Persistent Playgrounds
Events
• strengthen– reduces to refute and the strengthened claim has
Bob as defender– if strengthening not successful we have a
successful defense of the original claim• defender• opposer• claim• defended
05/03/23 16Persistent Playgrounds
Events
• agree– reduces to refute and Bob becomes also a
proposer of the claim
05/03/23 17Persistent Playgrounds
Old
05/03/23 Persistent Playgrounds 18
Why reputation
• if r1>r2: – claim C is defended by a proposer with against an
opposer with reputation r1 counts more than – claim C is defended against an opposer with
reputation r2 – difference reputation(proposer)-
reputation(opposer)
05/03/23 19Persistent Playgrounds
Comparison
New• tournament• second opinion evaluation• data mining
– strongly defended
Old• tournament• nothing• data mining
– two agree
05/03/23 20Persistent Playgrounds
Playground Comparisons
HSR• Kind: ForAllExists• Instance: (n,k)• InstanceSet: singleton• Solution: decision tree• quality: depth of tree• solve: minimize• claim: HSR(n,k)<=q• strengthen: yes, smaller q.• skills needed: Pascal’s triangle,
memoization
MMG• ForAllExists• x• [0,1]• y • f(x,y)=xy+(1-x)(1-y2)• maximize• k(c): AxEy f(x,y)>=c• yes, larger c.• skills: calculus, exploring a
surface
05/03/23 21Persistent Playgrounds