Reading Agendas Between the Lines, an exercise

Reading Agendas
between the lines
an exercise
Giovanni Sileno (g.sileno@uva.nl),
Alexander Boer, Tom van Engers
Leibniz Center for Law
University of Amsterdam
30 August 2016 - AI4J Workshop / ECAI 2016 @ Den Haag

social
system
How people
behave..

sources
of law
social
system
What the law
states..
How people
behave..

sources
of law
social
system
services
What the law
states..
How people
behave..
How public administrations
implements the law..

sources
of law
social
system
services
What the law
states..
How people
behave..
are three matters only loosely coupled

sources
of law
social
system
services
Focus on services/social system
What the law
states..
How people
behave..
●
non-compliance

sources
of law
social
system
services
Focus on services/social system
What the law
states..
How people
behave..
●
non-compliance
●
new user requirements

Diagnosis
●
Diagnosis starts from the presumption that a
failure occurred in the system.
●
But what counts as a failure?

Diagnosis of a designed artifact
●
In case of a designed artifact, we know the
function of the system, so if does not behave how
it was supposed to, this is a failure.

Diagnosis of a designed artifact
●
In case of a designed artifact, we know the
function of the system, so if does not behave how
it was supposed to, this is a failure.
●
Two types of failure:
– operational failure
– bad design

Consistency-based diagnosis
●
In case of electronic
components, for instance,
manufacturers furnish the
normal behaviour of the item.

●
failure = inconsistency with nominal specifications

●
●
Usual diagnostic problem:
recognize the minimal set of
components that produces the
inconsistency

●
-a :- not a.
In AI terms ~ default negation
implies strong negation

Abductive diagnosis
●
In other domains, we naturally create models of
faulty behaviour, because there may be non-
pathological cases in which things do not go as
expected. i.e. not all misalignments to the norm
are failures.

Abductive diagnosis
●
In other domains, we naturally create models of
faulty behaviour, because there may be non-
pathological cases in which things do not go as
expected. i.e. not all misalignments to the norm
are failures.
failure = consistency with explicitly faulty model
In AI terms ~ default negation is different from strong
negation

Diagnosis of a social system
●
What counts as a failure in a social system?

●
●
A wrong supposedly occurs when there are unmet
social expectations (for at least one of the
participants)

●
●
participants)
●
These are specified via normative directives:
prohibitions, obligations, etc.
– In principle, a wrong is detectable from the
violation of normative directives.

●
●
participants)
●
These are specified via normative directives:
prohibitions, obligations, etc.
– In principle, a wrong is detectable from the
violation of normative directives..
but proceeding “literally” is not sufficient.

Types of failure
e.g. duty to pay taxes

Types of failure
●
not to pay taxes (syntaxic, qualitative failure)

Types of failure
●
●
pay a wrong amount in respect to the declared
income (syntaxic, quantitative failure)

Types of failure
●
●
●
pay the correct but not the right amount, by using
tax evasion schemes (semantic failure)

Types of failure
●
●
●
The “semantic” failure is the
most difficult to be catched!

sources
of law
legal norms
Sources of knowledge

sources
of law
services
legal norms
business
process
models

sources
of law
services
legal norms
business
process
models
Usual focus of
compliance-
checking
methods

sources
of law
services
legal norms
business
process
models
Usual focus of
compliance-
checking
methods
..but not adequate to capture
the social semantics

sources
of law
social
system
services
stories,
experiences
legal norms,
cases
business
process
models

sources
of law
social
system
services
stories,
experiences
legal norms,
cases
business
process
models
Which representational
ground should we
consider?

sources
of law
social
system
services
agent-roles
stories,
experiences
legal norms,
cases
agent-role
models
business
process
models
Integrating model: agent-role

sources
of law
social
system
services
agent-roles
stories,
experiences
legal norms,
cases
agent-role
models
business
process
models
agent-role =
coordination of
roles (script +
topology)

sources
of law
social
system
services
agent-roles
stories,
experiences
legal norms,
cases
agent-role
models
business
process
models
agent-role =
coordination of
roles (script +
topology)
NB: the model abstracts the individuals

Domain: Real-estate transactions

Normal sale
Real estate A, worth 500.000 €

Normal sale
Real estate A, worth 500.000 €
With 6% transfer tax, if sold 30.000 € taxes

Hidden payment scenario
X wants to give to Y 300.000 €.
X sells Real estate A (worth 500.000 €) for 200.000 € to Y
with 6% transfer tax, 10.000 €.
(taxes for a direct transfer would be higher)

Swap-scheme scenario
Owner: X. Real estate A,
worth 10.000.000 €
Owner: Y. Real estate B,
worth 10.000.000 €
X wants B, Y wants A.
6% transfer tax: if sold,
1.200.000 € (total)

worth 10.000.000 €
worth 10.000.000 €
“Why don't we decrease
the nominal price?“

sold for 5.000.000 €
sold for 5.000.000 €
6% transfer tax:
600.000 € (total)

Swap-scheme topology
internal topologies – intentional coordination

internal topologies – intentional coordination
base for concentration of
interests (company,
family, etc.)

external topology – scenario coordination

external topology – scenario coordination
This may be contingent...

Taking the diagnoser view
bounded perception
●
Sale declarations
●
Sale payments

Monitoring and diagnosis
●
The observation abilities of the diagnoser are
necessarily limited.
●
Cognitive resources for reasoning are also limited.
It does not make sense to interpret all the data.

Monitoring and diagnosis
●
We consider an architecture affine to the Dual
Process theory of reasoning:
– monitoring (fast, reactive) for selecting
suspicious cases
– diagnosis (slow, reflective) for investigating
them.

Monitoring
Monitoring filters can be constructed by constrast
of normal cases with abnormal ones.
●
– Use a timeout rule, asynchronous check

Monitoring
●
– Use a timeout rule, asynchronous check
●
– Use an operational rule, synchronous check

Monitoring
●
– contrast determines threshold conditions

Monitoring
●
– contrast determines threshold conditions
In our case, adequate
thresholds are decided upon
average market price.

Selection rule in Prolog
suspiciousPrice(Price, Estate, Time) :-
marketPrice(MarketPrice, Estate, Time),
Price =< (MarketPrice * 60)/100.
suspiciousSale(Seller, Buyer, Estate, Price, Time) :-
declaration(sale(Seller, Buyer, Estate, Price, Time)),
suspiciousPrice(Price, Estate, Time).
Suppose the threshold is placed at -40% of market
price (e.g. after statistical analysis):

Diagnosis (main action)
ActionEvidenceOfSwap(
sale(Seller1, Buyer1, EstateA, PriceA, T1),
sale(Seller2, Buyer2, EstateB, PriceB, T2)
) :-
suspiciousSale(Seller1, Buyer1, EstateA, PriceA, T1),
suspiciousSale(Seller2, Buyer2, EstateB, PriceB, T2),
not(EstateA = EstateB),
not(Seller1 = Seller2), not(Buyer1 = Buyer2).
At this point, we check if suspicious sales are
consistent with a swap-scheme agent-role:

Diagnosis (additional evidence)
actionAndCircumstantialEvidenceOfSwap(
sale(Seller1, Buyer1, EstateA, PriceA, T1),
sale(Seller2, Buyer2, EstateB, PriceB, T2)
) :-
actionEvidenceOfSwap(sale(Seller1, Buyer1, EstateA,
PriceA, T1),
sale(Seller2, Buyer2, EstateB, PriceB, T2)),
relatedTo(Seller1, SharedStructure1),
relatedTo(Buyer2, SharedStructure1),
relatedTo(Seller2, SharedStructure2),
relatedTo(Buyer1, SharedStructure2).
..and when suspicious coordinated behaviour is found,
we look for circumstantial evidence about structures
of concentration of interests:

Swap-scheme generalization
●
Through computational means, we can
easily generalize the investigation to
n-steps swaps (obfuscasting
traces of coordination).
…this includes the
“hidden payment”
scenario!

Conclusion
●
This exercise showed a simplified application of a
more general framework

Conclusion
●
– The most important message is that...

Conclusion
●
agency trascends
individuals.

Conclusion
●
agency trascends
individuals.
..and we require abstractions like agent-roles
to pass to higher-level interpretation.

Further developments
●
– Prolog conflates strong negation with default
negation. What if some information is simply
unavailable? Other approaches (EBA) would be
more appropriate.

●
– Prolog conflates strong negation with default
negation. What if some information is simply
unavailable? Other approaches (EBA) would be
more appropriate.
– Scenarios should be collected in an adequate
representation for the modeler taking
causation into account. We suggested an
extension to Petri nets, with normative and
agentive positions.

●
– For the acquisition, in previous work we
proposed an intecremental method, starting
from UML-diagrams (similarly to requirement
engineering practices)

●
– For the acquisition, in previous work we
proposed an intecremental method, starting
from UML-diagrams (similarly to requirement
engineering practices)
– About computational complexity, existing
decomposition techniques could be used in
principle to compile the database of scenarios
offline, and for realtime exploitation.

Reading Agendas Between the Lines, an exercise

More Related Content

Similar to Reading Agendas Between the Lines, an exercise (20)

More from Giovanni Sileno (16)

Recently uploaded (20)

Reading Agendas Between the Lines, an exercise