1.
Introduction ^
2.
Background ^
2.1.
Related accounts in computer science & law ^
2.2.
The diagnosis and monitoring problem ^
The compliance/conformance decision making context is a diagnosis problem, and the associated information collection problem is a monitoring problem [4]. From the point of view of the compliance monitor, the objects of analysis are noncompliance stories. Abstract storylines are associated with a legal qualification backed by a legal rule: for instance client nonconformance, taxable event, internal performance problem, undecidable case, etc. A storyline consists of (a sequence of) events that happened, involving a number of agents that sent messages to each other, caused by, and resulting in, mental processes. Agent role behaviour scripts add the plan operators that causally connect the events, supplying both a rationale for the events that happened, and a basis for allocation of responsibility for the qualified events to specific agents. This responsibility assignment is a diagnosis in a narrow sense. An observation is essentially a received message by the diagnostic agent, caused by an action of the diagnostic agent. A sequence of observations may be consistent with a storyline, but should be logically distinguished from the events that make up the story, in which the diagnostic agent (usually) plays no role.
A diagnosis is a hypothesis that a subset of the agent roles is abnormal and the rest normal. As we argued in [4], default reasoning about agent behaviour can be modelled with a normal default theory [9]. For this default theory, Reiter’s default logic extensions are those of the generic diagnosis problem directed towards a minimal set of abnormal components [9]. Reiter essentially treats model-based diagnosis as a simple parametric design framework, where the main object is to allocate the qualifications {healthy, faulty} to components in a structure, meeting the minimality requirement.
The logic guiding the selection of actions that may expand the set of observations is complex and problematic. A number of factors plays a role in selection of the action:
- the diagnostic value of information,
- the reliability of the information that can be obtained,
- the costs of the action to obtain it, and
- the effects of obtaining it on future action.
3.
The problem space illustrated by some examples ^
This is a familiar storyline: the pickpocket’s diversion. The story is consistent with a number of very different hypotheses, and police officers will have a hard time producing evidence:
- The pickpocket used a spontaneously arising diversion to steal the bag, making the pickpocket the sole offender.
- An accomplice of the pickpocket picked a fight on the platform to create a diversion, allowing the pickpocket to steal the bag. The pickpocket and one of the two men involved in the fight are co-offenders.
- Two accomplices of the pickpocket simulated a fight on the platform to create a diversion, allowing the pickpocket to steal the bag. The pickpocket and both men involved in the fight are co-offenders.
- The passenger, guessing how the fight will be interpreted by the train conductor, uses the opportunity to make the train conductor a witness to the simulated theft of a non-existent laptop, in order to make an insurance claim. If the train conductor calls the railway police, the insurance claim of the passenger gains credibility. The passenger is the sole offender.
- The passenger left it on the railway platform. No offense took place. Perhaps the laptop bag was brought to a lost and found desk.
Now consider this more relevant storyline:
The cadastral register claims (e1) that a project developer sold an apartment for €350,000 to a natural person (e2). The payroll tax register claims (e3) that the project developer paid wages to that person (e4). The cadastral register claims (e5) that this person sold the apartment three months later for €500,000 to a third person (e6), but NVM real estate brokers claims (e7) that median home price hardly changed in that period (e8).
The reader of this paper, having his imagination already activated by the previous example, will undoubtedly realize that the information presented permits for many detailed hypotheses of varying plausibility, for instance:
- NVM median home price development is not representative for the specific property.
- The project developer, in order to avoid immediate insolvency, sold a property in great haste to one of its employees.
- The employee, acting as the agent of his employer, stole €150,000.
- The employer paid the employee €150,000, avoiding payroll taxes.
- The third person was forced at gunpoint to acquire a property worth €350,000 for €500,000.
- The employee, acting as the agent of his employer, was forced at gunpoint to cash in €150,000 from his employer, and an extortionist made of with the money.
- Another employee, who owed the first €150,000 euro, acting as agent of the employer, sold his property to the employer first, in order to extend the transaction chain and escape detection, and then to the other employee, who cashed in the difference, to evade income taxes.
- The project developer was a victim of bid rigging in a foreclosure auction, coordinated by the employee.
The events in the story are the result of the plans of the agents involved. Rather than understanding the story directly on the level of plans, we propose to explain the story on the level of agent role behaviour descriptions for the buyer and seller. This allows us to group together in a coherent unit the goals and adaptive plan operators that characterize
- plans for buying or selling, and
- plans for dealing with information seeking by third parties.
Let’s consider the example story in this context. If it is a simple case of tax evasion, the buyer and seller collude against the tax administration, and will coordinate testimony. They have reason to fabricate evidence for circumstances that would make the transaction normal, for instance by having the property retroactively appraised for €350,000 by a fraudulent professional appraiser, claiming great urgency, or claiming that the sale was agreed long ago when the reasonable market value of the property was €350,000. In the case of a theft, on the other hand, an agent representing the seller and the buyer colluded, but another agent representing the seller would likely offer reliable testimony. It is arguable that the agents were in a position to know that the suspicious transaction was going to be easily noticeable, and that therefore something more complicated, involving for instance extortion, could be going on. A thief or extortionist, becoming aware of having aroused suspicion, may react by threatening potential witnesses. Etcetera. These processes all influence the effectiveness of the monitoring process.
The following two behaviour descriptions for tax evasion satisfactorily explain event e2. e4 is only potentially relevant as a background to the motive for paying the employee. Plan operators follow the general pattern event (given conditions) ← plan, and adopted goals (intentions) in events and conditions start with to. First we present a script for the seller s:
- Goal: to pay an untaxed amount a to b.
- To pay an untaxed amount a to b ← propose to sell a property worth some v for v – a to b pretending the agreement was made long ago.
- b accepts proposal to sell a property worth some v for v – a←propose a specific property worth v to b.
- b accepts proposal of a specific property being worth v ← secure property for v then sell property for v – a to b.
- To sell property for v – a to b← offer to sell property for v – a to b.
- b accepts offer to sell property for v – a ← register property transfer to b in cadastral registration then monitor payment of v − a by b.
- Someone asks about the price v – a given the goal to pretend the agreement was made long ago ← claim the agreement was made verbally long ago.
In the interest of simplicity the plan is rather static. Note (2, 7) that fabrication of evidence plays a role in the planning. Next we present a script for the buyer b:
- Goal: to receive an untaxed amount a from s.
- s proposes to sell a property worth some v for v − a pretending the agreement was made long ago given the goal to receive an untaxed amount a from s ← accept proposal from s to sell a property worth some v for v − a.
- s proposes a specific property worth v to b ← have property for v appraised then consider proposal of specific property for v.
- Property for v was appraised given the goal to consider proposal of specific property for v and given that the appraisal shows that the property for v is acceptable ← accept proposal of the specific property being worth v.
- s offers to sell property for v − a to b given accepted proposal of a specific property being worth v ← accept offer to sell property for v − a then monitor registration of property transfer by s in cadastral registration then pay v − a to s.
- Someone asks about the price v − a given the goal to pretend the agreement was made long ago← claim the agreement was made verbally long ago.
- Someone asks about the appraisal given the goal to pretend the agreement was made long ago ← claim the agreement was an option to purchase, and the appraisal was made to decide about exercising the option then quickly inform s about the optional nature of the agreement.
Here again fabrication of evidence plays a role (2, 6, 7). In this plan, having the appraisal performed may attract attention: why have the property appraised before acceptance if the agreement already exists? Step 7 is a plausible, but – from a timing point of view – hairy improvisation in response to the monitor, that must be supported by s.
The associated monitoring script interacts with the cadastral registry, payroll data registry, online NVM database, appraisers, buyers, and sellers. It is omitted for reasons of space, considering its complexity even in a simplified domain model. Let us assume the investigation completes the story with the following interpretation of events by the monitor:
The parties involved agree (e9, e10, e11) that the reasonable market price is €500,000 (e12). The project developer (e13) and its employee (e14) both claim to have agreed previously to an option to purchase the property for a then reasonable price (e15). The employee later also claimed (e16) that the agreement was an option to purchase, and the appraisal was made to decide about exercising the option (e17). The tax administration claims (e18) that the verbal agreement claimed in e15 is an unverifiable result of a fabrication for the purpose of evading taxes (e19). The project developer and employee countered with the claim (e20, e21) that the relationship of trust that exists between them made requiring a written agreement imprudent (e22). The tax administration claims (e23) that the difference between the reasonable price and the actual price paid (€150,000) is taxable income (e24).
From an explanatory point of view, e24 is the conclusion, supported by e12 (e9, e10, e11) and e2 (e1). The argument for a contrary conclusion is supported by e15 (e13, e14). If e19, the weak attack in the tax administration’s reconstruction on e15, is not accepted by the tax court, a new policy guideline points the way out: only written options to purchase real estate should be enforceable, and they should be duly registered somewhere to prevent antedating. More generally, the tax administration’s explanation of events is that the project developer and employee were following scripts s and b, and that the tax administration’s specific monitoring policy in this case, script m, and the opportunities created by the civil code’s rules on agreements, which require that real estate transactions must be provable to third parties, but options to purchase need not be, makes enforcement of tax evasion a problem in this case. A note of caution: this plan is obviously not state-of-the-art in tax evasion.
4.
A knowledge acquisition framework for compliance controls ^
In [3] we discussed the use of AgentSpeak scripts representing typical behaviours, and noted that finding a logical organization of such scripts is an open problem. In this presentation, which is about inference, we structure scripts in such a way that it is easy to present specific parametric design and monitoring problems in an ontology as an assignment of instances to classes. The generic components of the knowledge acquisition framework are the following:
- A set of agent role instances ARI = {ari1, ari2, . . . , arin}
- A set of agent role classes ARC = {arc1, arc2, . . . , arcn}
- A complete 1:n relation arc(ari) of agent role classes to instances (for all ari ∈ ARI there is an (ari, arc) ∈ ARIC where arc ∈ ARC)
- A set of normal behaviour scripts for each agent role class, NS(arc) = {ns1, ns2, . . . , nsn}
- A set of abnormal behaviour scripts for each agent role class, ANS(arc) = {ans1<, ans2, . . .., ansn}
- A sequence of events E = {e1, e2, . . . , en}, where events are messages m(a1, a2, p), as given before in section 2, and p may again be a proposition about a message.
- Some arbitrarily structured background theory BT that permits consistency checking.
An agent role behaviour script allocation is an n:1 relation ALLOC, where for all (ari, ns) ∈ ALLOC it is the case that ari ∈ ARI, arc(ari) ∈ ARIC, and either ns ∈ NS(arc) or ns ∈ ANS(arc). A complete allocation allocates a script to all ari ∈ ARI. An allocation is an explanation of E, given BT, if E follows from ALLOC∪BT, ALLOC does not already follow from BT ∪ E, and ALLOC ∪ BT ∪ E is consistent (cf. for instance [5] for compatible accounts). Explanations need not be complete, and a partial allocation may form part of the problem definition.
A diagnostic explanation of E is an allocation ALLOC, where, for at least one (ari, ns), ns ∈ ANS(arc). This does not necessarily mean that all non-diagnostic explanations do not fit. A diagnosis suggests a repair: a diagnostic explanation leads to a legally or administratively relevant and actionable qualification of the behaviour. This is a problem domain in itself, not addressed here. That there is only a diagnosis problem if all non-diagnostic explanations have been rejected is a simplifying assumption in problem solving frameworks, not a reflection of reality. The ultimate objective is not to reach an unambiguous conclusion, but simply to select a next action.
Actual design and explanation problems, as dealt with in public administration, depend on splitting the set of agent role instances into several subsets. There is the monitoring system (MON ⊂ ARI) that is monitoring, and the domain (DOM ⊂ ARI\MON) being subjected to monitoring. In the domain DOM in public administration we find clients and network partners outside (OUT ⊂ DOM, NET ⊂ OUT, CLIENTS ⊂ OUT), and we find case handlers on the work floor inside the organization (IN ⊂ DOM). Compliance monitoring seeks an explanation for CLIENTS, or OUT, or DOM, while other script allocations are part of the background theory.
5.
Discussion ^
6.
References ^
[1] Verschueren, J.: The impact of legislation: a critical analysis of ex ante evaluation. Martinus Nijhoff publishers (2009)
[2] Gribnau, J.L.M., Lubbers, A.O., Vording, H.: Terugkoppeling in het Belastingrecht. SDU (2008)
[3] Boer, A., van Engers, T.: An agent-based legal knowledge acquisition methodology for agile public administration. In Ashley, K.D., van Engers, T.M., eds.: The 13th International Conference on Artificial Intelligence and Law, Proceedings of the Conference, June 6–10, 2011, Pittsburgh, PA, USA, ACM (2011)
[4] Boer, A., van Engers, T.: Diagnosis of multi-agent systems and its application to public administration. In Abramowicz, W., ed.: Business Information Systems Workshops – 14th International Conference, BIS 2011, Poznan, Poland, June 15–17, 2011. Volume 97 of Lecture Notes in Business Information Processing, to appear., Springer (2011)
[5] Lipton, P.: Inference to the best explanation. Philosophical issues in science. Routledge (1991)
[6] Keppens, J., Schafer, B.: Knowledge based crime scenario modelling. Expert Syst. Appl. 30(2) (2006) 203–222
[7] Bex, F.J., van Koppen, P.J., Prakken, H., Verheij, B.: A hybrid formal theory of arguments, stories and criminal evidence. Artif. Intell. Law 18(2) (2010) 123–152
[8] Hoekstra, R., Breuker, J.: Commonsense causal explanation in a legal domain. Artif. Intell. Law 15 (January 2007) 281–299
[9] Reiter, R.: A theory of diagnosis from first principles. Artificial Intelligence 32(1) (1987) 57–95
[10] Howard, R.: Information Value Theory. IEEE Transactions on Systems Science and Cybernetics 2(1) (1966) 22–26
[11] Motta, E., Zdrahal, Z.: Parametric design problem solving. In: Proceedings of the 10th Banff Knowledge AcquisitionWorkshop. Number 1990, SRDG Publications, University of Calgary (1996) 1–24
Alexander Boer. Leibniz Center for Law, University of Amsterdam, The Netherlands. E-mail: A.W.F.Boer@uva.nl.
Tom Van Engers. Leibniz Center for Law, University of Amsterdam, The Netherlands
This article is republished with permission of IOS Press, the authors, and JURIX, Legal Knowledge and Information Systems from: Kathie M. Atkinson (ed.), Legal Knowledge Systems and Information Systems, JURIX 2011: The Twenty-Fourth Annual Conference, IOS Press, Amsterdam et al.