A Constructivist Approach to Rule Bases
Giovanni Sileno, Alexander Boer and Tom Van Engers
Leibniz Center for Law, University of Amsterdam, Amsterdam, The Netherlands
Keywords:
Rule Bases, Constructivism, Rules Revision, Rules Optimization, Adaptation, Regulated Systems.
Abstract:
The paper presents a set of algorithms for the conversion of rule bases between priority-based and constraint-
based representations. Inspired by research in precedential reasoning in law, such algorithms can be used for
the analysis of a rule base, and for the study of the impact of the introduction of new rules. In addition, the
paper explores an optimization mechanism, built upon assumptions about the world in which the rule-based
system operates, providing a model of environmental adaptation. The investigation is relevant to practical
reasoning, agent modeling and agent programming.
1 INTRODUCTION
Practical reasoning in everyday life is often not based
on repeated deliberation, but on behavioural scripts
constructed and refined by some adaptation process.
Let us consider this simple story: Raphael, who lives
in a rainy country, always checks before he goes to
work whether it will rain, in which case he will bring
his umbrella. Samuel, who lives in a sunny country,
usually goes out without checking the weather. Both
of them believe in taking an umbrella if it rains, how-
ever. Intuitively, if Raphael or Samuel moved from
one country to the other, we would expect that they
would change their habits eventually.
According to traditional optimization theory,
adaptation comes from the agent’s efforts to obtain
a better overall pay-off. In contrast, Heiner’s theory
of predictable behaviour (Heiner, 1983) explains how
behavioural regularities arise in the presence of unre-
solvable uncertainty about the “right” course of action
to follow: being predictable seems to pay off by itself.
Regardless of its explanation, we acknowledge the
existence of a sort of structuring process in the agent,
which results in behavioural rules or scripts that are
followed without consciously reflecting on them. The
foundations of agent modeling, agent programming,
and before those, of expert systems, are for the most
based on such behavioural rules or scripts.
A similar analysis can be performed on collective
agencies, whose behaviour is described/prescribed
via formalized artifacts and procedures. Focusing on
administrative organizations, Boer and Van Engers
recognize in (Boer and Engers, 2013) three spheres of
activity: operations, planning/design, policy making.
The three spheres are concerned by different activi-
ties and often utilize different resources. For instance,
people at the operations provide services depending
on the procedures and the resources assigned by the
design department.
For simplicity reasons, this paper divides the
agency/system in only two levels: the regulated (non-
reflective) component and the regulatory (reflective)
component. In contrast to approaches like machine
learning, theory induction, etc. our agent is not con-
structing the rules that govern his behaviour from the
facts he observes, but he follows the prescription a
given set of rules (cf. agent programming), possibly
making its operationalization more efficient.
Two research questions are mainly addressed.
First, how to revise (the operational knowledge of) a
rule base if a new rule is introduced? This problem is
common in practical reasoning, statutory law and case
law. Horty recently explored in (Horty, 2011) the dy-
namics of normative systems based on common law,
examining phenomena like distinguishing on binding
precedent rules. Inspired by his work, we generalize
the analysis of revision in law to a generic intelligent
system whose operational knowledge is described by
a rule base. Focusing on two ways to represent a
system of rules, priority-based and constraint-based
2), we identify and organize the algorithms to pass
from one representation to the other (§ 3).
Then, analyzing the interactions of the evaluation
of rules with default assumptions, we extend the anal-
ysis to consider the rewriting of the rule base due to
new informational commitments, in order to tackle
540
Sileno G., Boer A. and van Engers T..
A Constructivist Approach to Rule Bases.
DOI: 10.5220/0005280205400547
In Proceedings of the International Conference on Agents and Artificial Intelligence (ICAART-2015), pages 540-547
ISBN: 978-989-758-074-1
Copyright
c
2015 SCITEPRESS (Science and Technology Publications, Lda.)
down the second question: how to refine rules when
certain expectations are ascribed to the environment?
4) As the change of informational commitment and
the beginning of the rule revision process are only
asynchronously linked, the coupling is worth further
analysis (§5).
Both the rules that govern the practical behaviour
of the agent and the default assumptions about the
world may change. For these reasons, our work is
affine to the spirit of defeasible logic. Rather than
focusing on a formal framework following this tra-
dition, however, we prefer at this stage to propose a
direct computational implementation.
2 REPRESENTATION TYPES
We summarize some practical definitions that will be
used throughout the paper.
A fact f is a true proposition.
A rule r is a conditional statement relating a
premise with a conclusion. Premise and conclu-
sion are propositions, or expressions of proposi-
tions. They are called also body and head.
Given a rule r, Premise(r) is a function returning
the set of atomic propositions that should be true
when its premise is true. These components are
also called factors. Conclusion(r) is a function
returning the atomic propositions asserted by the
rule when the premise is true.
A situation Σ is a set of facts: Σ = { f
1
, f
2
, . . . , f
n
}.
The applicability condition is expressed by Σ |=
Premise(r).
A rule base is a set of rules: = {r
1
, r
2
, . . . , r
n
},
possibly partially ordered.
Evidently, different rules may share part of the
premises, or stated equivalently, share part of their
domain of applicability. In general, a set of facts can
apply on several rules. Furthermore, when attacking
each other conclusions, rules are potentially in con-
flict. When conflicting rules apply, we have the prob-
lem of determiningwhich is the right conclusion. Two
solutions are possible:
at horizontal level, modifying adequately the con-
tent of rules, i.e. redefining the constraints of ap-
plicability of mutually conflicting rules;
at vertical level, defining relations of hierarchy or
priority between rules.
Two representational attitudes are directly related
to those approaches: constraint-based and priority-
based.
2.1 Constraint-based
A rule base written following a strict constraint-based
approach is a non-ordered set. The relative position
of rules (e.g. the position in the artifact describing the
regulatory system, the chronological time of creation,
etc.) is not relevant. A rule is required to transport all
what is needed to check its applicability, i.e. all the
relevant factors have to be made explicit.
On the down side, however, the complete set of
rules has to be (maintained) consistent as a whole.
The problem can be explained referring to the pro-
gramming construct switch...case, which separates
execution paths depending on a certain condition:
switch (condition) {
case value : [...]; break;
[...]
}
In order to be functional, the condition variable has to
index precisely one value position in the list.
As the application of a rule is triggered by the rel-
evant factors in the premise being true, the most basic
encoding is built upon the permutations of all factors
described in the rule base. If the rule base is sensitive
to N factors, there are 2
N
situations to be encoded.
Such full-tabular encoding implies that the requests
for a rule are inflected also in terms of factors which
are irrelevant for the rule in itself, but are relevant to
other rules. Other more efficient encoding are possi-
ble.
2.2 Priority-based
In all acts of communication there is an intrinsic
strictly ordered dimension, usually referred to as the
discourse. It can be temporal, as in the case of inter-
actions; or positional, as with texts.
Priority-based representations take advantage of
the meta-information intrinsic to the composition of
content to reduce its redundancy. In programming,
something similar occurs with the construct if...else:
if (condition1) [...];
else if (condition2) [...];
else [...];
In order to avoid to repeat the same calculations, a
good programmer implements the conditions to be
checked in a convenientway, conscious of the sequen-
tial processing of the code. For instance, he knows
that if condition2 is being evaluated, condition1 is
necessarily false.
Considering such structured enchaining in the
evaluation of rules, each new evaluation could remove
the conjunction of factors which have been already
AConstructivistApproachtoRuleBases
541
computed. The resulting rewriting of the rules is more
efficient(in size terms) than the original version. Con-
versely, the application of a rule may require more
computational effort than a constraint-basedrepresen-
tation, as the applicability must be strictly checked
following the given ordering.
So far, we have considered priority as directly re-
lated to ordering (positional or chronological). In
such conditions, if we introduce a new rule with a cer-
tain domain of application, we overwrite all existing
rules whose domains subsume that of the new rule.
There are situations in which it is useful to avoid this
mechanism. This occurs for instance in law with the
principle called lex specialis derogat legi generali: a
more specific law overrules the more general. Princi-
ples like this create meta-rules: their conclusion is not
a fact, but a priority constraint. In the following we
suppose that we are always able to apply meta-rules
to obtain the relative priority between rules.
1
3 INCREMENTAL DYNAMICS
Depending on the underlying representation, inter-
ventions on a rule base require a different attention:
in a constraint-based rule-base, we need some
kind of rule revision process (Horty calls it refine-
ment) on the rules whose applicability is affected
by the addition, so that after the update each rule
is again consistent with the whole, and can be also
read separately;
in a priority-based rule-base, the addition of rules
or modification of priorities is straightforward;
however, we cannot infer the resulting practi-
cal consequences without taking into account the
whole rule-base.
Our objective is therefore to provide a conversion
method from one representation to the other, in order
to cope with the respective limitations. While the nat-
ural way of handling the incremental construction of
a rule-base is via priority-based representations, the
more direct way to access the actual applicability of a
rule comes from a constraint-based representation.
From Compound to Simple Rules. In the follow-
ing, we will often refer to simple rules, i.e. rules
with no disjunction in the body, nor conjunction in
the head. They can be seen as the atomic components
1
For instance, the implicit ordering of rules with fac-
tors expressed at different levels of generalization can be
obtained via Pearl’s system Z (Pearl, 1990).
of the conditional. Compound rules can be reduced to
simple rules applying the following properties:
(a c d) (a c) (a d) (1)
(a b c) (a c) (b c) (2)
Relevance. Given two simple rules r
1
and r
2
, we
say that r
1
is relevant in respect to r
2
, if the propo-
sitional content of their conclusions are positively or
negatively aligned: Conclusion(r
1
) = Conclusion(r
2
)
or Conclusion(r
1
) = ¬Conclusion(r
2
).
2
Illustrative Example. Let us assume the existence
of a principle like lex posterior derogat legi prior,
i.e. the last rule has higher priority than the previ-
ous ones. Consider a rule base consisting of just one
rule r
1
: A P. With one factor (A), we can have two
situations, but our rule refers only to one of them. The
associated representations are:
3
Constraint-Based
Priority-Based
A P A P
[¬A ?]
At this point, we add a second rule: r
2
: B ¬P.
Considering two factors (A, B), we have four possi-
ble situations. In a constraint-based representation,
we have to refine the previous rule in order not to lose
consistency, as r
1
and r
2
have opposite conclusions,
and therefore may be in conflict. Removing the do-
main of applicability of r
2
from the domain of r
1
we
obtain the following representations:
CB (full-tabular)
CB (intermediate) PB
A ¬B P A ¬B P A P
A B ¬P B ¬P B ¬P
¬A B ¬P
[¬A ¬B ?]
The table distinguishes two types of constraint-
based representations: full-tabular, where the con-
figuration of all factors is written explicitly for each
rule, and intermediate, where redundancy is reduced
via boolean simplification.
We follow this illustrative example to propose
some algorithms to transform one representation into
another. This is sufficient to handle the problem of in-
cremental construction. Given a constraint-based rule
base , the refinement required for the introduction
2
This is a “narrow” definition of relevance. In general,
given a rule, we care also of the rules whose conclusions are
premise for the rule, etc.
3
For completeness, we use the expression ? to iden-
tify the lack of consequence, for situations which do not
imply any conclusion.
ICAART2015-InternationalConferenceonAgentsandArtificialIntelligence
542
of a new rule r
n
is equivalent to the conversion of the
priority-based rule-base +{r
n
}, where r
n
has higher
priority than all other rules.
3.1 From Priority-based to
Intermediate Constraint-based
Representation
Suppose we have a strictly ordered set of simple rules
= {r
1
, r
2
, .., r
n
}, where r
1
is the rule with lowest pri-
ority, r
n
with the highest. The following algorithm
overwrites the elements of with their translation in
a constraint-based intermediate representation:
4
for each r
i
, starting from r
n
down to r
2
for each r
j
, starting from r
i1
down to r
1
if Conclusion(r
j
) = ¬Conclusion(r
i
) then
r
j
:=
V
Premise(r
j
) ¬
V
Premise(r
i
)
Conclusion(r
j
)
Stated in words, each cycle adds the complement
of the domain of applicability of the rule with higher
priority to the rule with lower priority.
The negation in the core formula usually applies to
sequence of multiple and expressions, so it is equiv-
alent a sequence of or expressions. Using the body
disjunction formula (2) we can divide such compound
rule in its simple components, and some of those may
be redundant with other rules in the rule base. This is
not the only potential cause of redundancy: for histor-
ical or clarity reasons, redundant rules may have been
introduced already in the priority-based account (ex.
r
3
: A ¬B P). As a consequence, the previous
algorithm may produce subsumed or cloned rules
5
,
which can be removed without harm.
3.2 From Intermediate to Full-tabular
Constraint-based Representation
The same regulation can be described with several in-
termediate representations. A way to compare them
is to fully compute their extensional meaning, i.e.
the underlying full-tabular representation. In order to
do that, we can expand each rule with the unspeci-
fied factors which are relevant (i.e. are part of the
premises) for relevant rules (i.e. having the same con-
clusion or its complement) in the rule base.
For instance, in the illustrative example, r
1
and r
2
are relevant to each other, and refer to the two factors
4
V
{ f
1
, f
2
, ..., f
n
} f
1
f
2
... f
n
5
Given two rules r
i
and r
j
such that Conclusion(r
i
) =
Conclusion(r
j
), if Premise(r
i
) Premise(r
j
) then r
i
sub-
sumes r
j
; if Premise(r
i
) = Premise(r
j
) then r
i
clones r
j
.
A and B. However, r
2
: B ¬P does not explicitly
refer to the factor A, but it can be rewritten as two
full-specified rules using some boolean properties:
B ¬P (B A ¬P) (B ¬A ¬P)
If a rule has n unspecified factors in the constraint-
based intermediate representation, then it is written as
2
n
rules in the full-tabular representation.
3.3 From Full-tabular to Minimal
Constraint-based Representation
Our objective at this point is to find a way to produce
a minimal intermediate form, i.e. removing all redun-
dancy. The solution we propose is to pass by the full-
tabular representation and then simplify using known
algorithms for boolean simplification.
We have a set of rules = {r
1
, r
2
, .., r
n
}, repre-
sented in a full-tabular constraint-based way. Its char-
acterization can be visualized using a table similar to
a truth table. The only difference is that we are not
interested in showing the specific truth value of the
output variable for all combinations of inputs, but in
mapping where specific outputs are defined. Consid-
ering the illustrative example, we have:
A
B P ¬P ?
T T X
T F X
F T X
F F X
To reduce the implicit redundancy of the full-
tabular representation, we can apply then the same
principle of the minterm expansion used for logic cir-
cuits, in order to compute the characteristic function.
First, we aggregate all rules which share the same
conclusion φ into seed rules r
φ
:
for each r
i
, with r
i
6= r
φ
φ := Conclusion(r
i
)
if r
φ
is not set then r
φ
:= r
i
else r
φ
:=
V
Premise(r
φ
)
V
Premise(r
i
) φ
Second, we apply an algorithm for boolean reduc-
tion e.g. Quine-McCluskey (Mccluskey, 1956)
6
to reduce each r
φ
to its minimal form.
These algorithms take as input the logical addition
(disjuction, or) of the products (conjunction, and) of
factors producing a target φ. The output is a rule r
φ
which may be a compound rule. In this case, it can be
6
The Quine-McCluskey algorithm provides an optimal
solution, but is costly in computational terms. There are
heuristics which provide suboptimal solutions consuming
less computational power exist, e.g. Espresso.
AConstructivistApproachtoRuleBases
543
divided in its components r
φ1
, r
φ2
, ... using the body
disjunction formula.
Finally, considering both positive and negative
characterizations of a certain conclusion φ, we obtain
φ
= {r
φ1
, r
φ2
, . . . , r
¬φ1
, r
¬φ2
, . . .}, which is by con-
struction the minimal constraint-based representation
of the φ-relevant component of the initial rule-base.
3.4 From Constraint-based to
Priority-based Representation
Suppose we have a set of rules , given in a minimal
constraint-based representation. In order to obtain
an associated priority-based representation we need
some partial ordering between the rules. Such order-
ing can be: (a) hard-coded, i.e. provided as input,
(b) obtained by some evaluative function which takes
the rules as inputs (§ 4). In the following we suppose
the ordering is already available; for simplicity, we
assume that the input priorities constructs a strict or-
der on the rule set (we label r
n
the rule intended to be
with the highest priority, r
1
with the lowest one). The
following algorithm performs the transformation of
according to such priority labeling:
for each r
i
in , from r
n
to r
1
φ = Conclusion(r
i
)
factors := RelevantFactors(, φ)
if notYetEvalSituations
φ
is not set
notYetEvalSituations
φ
:= Allocate(factors)
if establishedFactors
φ
is not set
establishedFactors
φ
=
create two empty rule bases
i
and
i
add r
i
to
i
convert
i
to full tabular considering factors
for each r
j
i
if Premise(r
j
) notYetEvalSituations
φ
newPremise := Premise(r
j
); apply := true
for each f establishedFactors
φ
if ¬ f Premise(r
j
)
apply := false; break
else if f Premise(r
j
)
newPremise := newPremise\ { f}
if apply newPremise 6=
r
j
:=
V
newPremise φ; add r
j
to
i
notYetEvalSituations
φ
:=
notYetEvalSituations
φ
\ Premise(r
j
)
establishedFactors
φ
:=
ExtractFacts(notYetEvalSituations
φ
)
apply Quine-McCluskey on
i
obtaining r
i
The general principle of the algorithm is to cre-
ate clusters of rules depending on the conclusion, and
to synthesize the reciprocal dependency only within
these clusters. We leverage the following information:
which factors are relevant, considering the whole rule
base, in respect to the conclusion [Relevant(, φ)]; the
set of all possible situations, via the allocation of a
truth value to such relevant factors [Allocate(·)]. Ex-
panding the rule to its full-tabular components, we
prune the situations which are evaluated, and we ex-
tract the set of factors which have been established on
the not yet evaluated situations [ExtractFacts(·)].
7
4 INTEGRATING ASSUMPTIONS
In a constraint-based representations all relevant and
discriminatory factors are made explicit. In an opera-
tional setting, however, it is difficult and not efficient
to evaluate all of them: the check of conditions may
require external investigations.
Following optimization principles, it is important
to put in the beginning checks on conditions which
cost less (more in general, have better pay-off) and
provide more discrimination in the set of possible ex-
ecution paths (i.e. are more informative, in Shannon’s
terms). This means that knowledge about the world
and of the operational mechanisms of the system do
provide elements to create a priority in the evaluation.
Consider for instance this example:
if (f()) [...];
else if (g()) [...];
else [...];
This implementation would be really poor if
f()
is
very costly and often false, while
g()
is not computa-
tionally expensive and often true.
4.1 Payoff Analysis
In general, the expected payoff in performing an ac-
tion can be decomposed as:
E[payoff] = p(success) · E[payoff of success]
+ p(failure) · E[payoff of failure]
Actions are expected to have some side effects, which
may occur at symbolic level (e.g. acquiring new in-
formation) and at practical level (e.g. modifying the
physical environment). The second aspect may be
critical in certain settings. For instance, trying to get a
certain evidence often brings the risk to destroy other
evidence. Such side effects modify what the agent
could do in the following step. For simplicity, we ne-
glect this dynamic component in this work. Consider-
ing a (non invasive) action, the most important voice
7
An implementation in Java/Groovy of the algorithms
is available on our site. http://justinian.leibnizcenter.org/
rulebaseconverter
ICAART2015-InternationalConferenceonAgentsandArtificialIntelligence
544
related to side effects remains a quantification of the
expected cost.
Suppose that the action is an investigation. The
success corresponds to reach a certain conclusion
about φ. Assuming the cost is known and the same
for both success and failure, the previous equation be-
comes:
E[payoff] = p(success) · E[payoff of concluding φ]
+ p(failure) · E[payoff of non concluding φ]
cost
The investigation is worth if E[payoff] > 0, i.e. if
E[payoff of concluding φ] >
cost (1 p(success)) · E[payoff of non concl. φ]
p(success)
In the case in which no positive or negative out-
come is entailed from not concluding φ, the previous
constraint is simplified to:
E[payoff of concluding φ] >
cost
p(success)
(3)
4.2 Rule Guided Investigation
Suppose our knowledge is described in the form of a
rule base. Let us consider a target matter φ. We know
that the factors concerning φ, i.e. the factors rele-
vant for rules implying φ or ¬φ, are the reasons which
possibly allow us to conclude φ or ¬φ. In Shannon’s
terms, they provide the dictionary of symbols which
should be possibly extracted from the informational
source for our investigative purposes.
4.2.1 Individual Evaluation
Suppose we have a rule whose conclusion is φ, be-
longing to a rule base described in a constraint-
based representation. The probability that the state
of the world satisfies the domain of applicability of r
is given by:
p(applicable(r)) = p(
^
Premise(r))
An action on investigation performed following a cer-
tain rule is successful if it is able to bring to the in-
tended conclusion φ. Therefore, accounting the set of
known facts K:
p(success) = p(
^
Premise(r)|K)
Going further, the function quantifying the informa-
tional cost can be written in the form c(D, K), where
D is a set of factors we need to investigate, and K is
the set of known facts.
cost = c(D, K)
Thus, the rule r should be evaluated only if:
E[payoff of concluding φ] >
c(D, K)
p(
V
Premise(r)|K)
4.2.2 Optimization
In general, the rule base may contain several rules
which entail φ. The choice about which rule should
be evaluated first is optimal if the rule maximizes the
investigative payoff.
The payoff of reaching a conclusion about φ is in-
dependent of the rule used, therefore can be labeled
as a constant:
G = E[payoff of concluding φ]
Furthermore, in respect to a rule r, the set of desired
factors D consists of the factors in the premise which
are not known:
D = Premise(r) \ K
Neglecting the payoff of not concluding φ, we have:
E[payoff] = G· p(
^
Premise(r)|K) c(D, K) (4)
The next paragraphs consider a few simple configura-
tions.
Equal Probabilities. Suppose that there are n fac-
tors relevant for φ, statistically independent and with
equal probability. Supposing c(D, K) negligible, we
have:
E[payoff]
1
n
#D
Suppose also we are ignorant about the world, there-
fore #D = #Premise(r). In a full-tabular constraint-
based representation, all rules have the same num-
ber of premises, therefore the expected payoff is
the same for all rules. Conversely, in a minimal
constraint-based representation, rules have been con-
structed by functional aggregation reducing the num-
ber of premises. Therefore the optimal choice is to
choose the rule with less conditions in the premise.
Unequal Probabilities. In condition of ignorance,
and with statistically independent factors, we have
p(Premise(r)) = p( f
1
) · p( f
2
) · . . . · p( f
n
)
At this point, it is interesting to refer to the theory
of communication introduced by Shannon (Shannon,
1948). Defining the information related to a certain
symbol s as:
I(s) = log
2
(p(s))
the information of a message m composed of different
and statistically independent symbols (s
1
, s
2
, ..., s
n
)
is given by the sum of information of each symbol:
I(m) = I(s
1
) + I(s
2
) + . . . + I(s
n
)
AConstructivistApproachtoRuleBases
545
Thus, assigning each factor to a symbol, the premise
of a rule r can be interpreted as a message. As infor-
mation maintains monotonicity, we can consider the
best rule to be chosen not in terms of maximizing the
product of probabilities, but as minimizing the sum of
information required to reach the target conclusion,
possibly integrating the cost component.
4.3 Default Assumptions
Let us consider again the story reported in the intro-
duction. Both Samuel and Raphael think it is normal
to take the umbrella if it is raining. Simplifying
8
, we
can write this practical rule as:
rain umbrella
With no prior assumptions, the rule base of both
agents consists of a simple rule. Now, suppose the
agents have a probabilistic model of the world. Ac-
cording to (4), the payoff of the evaluation of that rule
is equal to:
E[payoff] = G· p(rain) c({rain}, K)
If the agent already knows the fact that it will rain,
supposing retrieval costs negligible, the overall cost c
is null. In such case, the payoff of the investigation is
always positive. It tends to 0 when the agent thinks
that the probability of rain is very low, as in the case
of Samuel, living in a sunny country: p(rain) 0.
However, if Samuel’s knowledge base does not
contain the “rain” fact, the expected payoff of the in-
vestigation may be negative: if G · p(rain) is small
enough, the payoff constraint is violated, so the rule
shouldn’t be evaluated.
9
Raphael is in the opposite
condition: as p(rain) is not null, the positive addend
in the expected payoff of investigation makes the cost
of checking the weather negligible.
4.3.1 Prioritization with Default Rules
The analysis of evaluation payoffs provides an op-
timal order of investigation and such ordering can
feed the conversion from constraint-based to priority-
based representations 3.4). However, if the ex-
pected payoff is null or negative, the related investi-
gation is not worth and the agent should rather con-
sider to introduce a default assumption rule. A simple
8
We put aside all issues related to causality, intentional-
ity, action, etc.
9
The payoff G of concluding the fact that it rains can be
interpreted in terms of the practical consequences that such
conclusion entails. For instance, the agent will avoid to be-
come wet, by taking the umbrella. How much this is valu-
able to the agent depends on subjective components, and
plausibly changes in time.
way to represent that is to introduce the negation as
failure operator, as defined for instance in Answer Set
Programming (Baral and Gelfond, 1994; Lifschitz,
2008). In practice, when we commit to a default as-
sumption concerning the fact f , we add to the rule
base the following rules:
(not¬ f f) (not¬ f not f ¬ f)
Note that there is an implicit priority between the
two rules, cf. system Z (Pearl, 1990). Such interven-
tion does not modify the previous rules, but overrides
the investigations concerning f. In other words, they
force the agent not to perform the investigation but
just to refer to his current knowledge, so that the pay-
off constraint is not violated.
10
5 CONSTRUCTION AND
RECONSTRUCTION
The paper investigated two types of events changing a
rule base. The first type of events consisted of incre-
mental modifications: new rules are added by external
intervention, determining a partial reconfiguration of
the operationalknowledgeused by the agent. Because
of distinguishing actions, the new rules brings to the
foreground factors left implicit in the previous rules.
The second family of events concerned ad-hoc re-
organizations, aiming for better adaptation. However,
when a rule base is “compiled to a more efficient
priority-based form, the agent loses the reasons moti-
vating that structure (e.g. probabilistic assumptions),
and therefore he cannot check if those reasons are still
valid.
To rewrite the rule base again, the agent has to re-
flect over the rule base. He has to unveil the underly-
ing constraint-based representation, removing all de-
fault assumptions and recompute the optimal priority-
based indexing.
What may motivate the agent to do that? An in-
tuitive answer is because of (repeated) failures of his
current practical reasoning. For instance, if Samuel
moves in a rainy place, his default assumption about
the rain will lead him to get wet often. If the num-
ber of practical failures exceeds a certain (subjective)
threshold, we expect the agent will look for a better
adaptation to the context, asynchronously starting the
reflection cycle.
10
There may be more radical transformations. For in-
stance, destructive simplifications can be imagined remov-
ing f from the premises of the rules, removing from the rule
base rules which have ¬ f in the premises, etc.
ICAART2015-InternationalConferenceonAgentsandArtificialIntelligence
546
6 CONCLUSION
Part of a wider research concerning agility in pub-
lic administration and policy making (Boer and En-
gers, 2013), the present paper starts the development
of a computational framework operationalizing a con-
structivist approach to rule bases. By dividing agency
in a regulated and a regulatory sub-systems, we ex-
plicitly disjoin the processing of facts, depending on
the rule base, from the modification of the rule base.
The analysis we presented is not targeting
beliefs—in the traditional sense of the belief revision
literature, e.g. (Ferm´e, 2011)—norbuilt upon a model
of dynamic theory revision of knowledge accounting
for both facts and rules, as in machine learning, see
e.g. (Omlin and Giles, 1996), (Goldsmith and Sloan,
2005). Similarly to Horty (Horty, 2011), our specific
scope is on rules, as components of a rule base, al-
ready defined at symbolic level.
In psychology, the theory of constructivism is tra-
ditionally related to Jean Piaget (Piaget et al., 2001),
who investigatedthe mechanisms under which knowl-
edge is internalized by learners. He argued that in-
dividuals construct their knowledge through the two
processes of:
assimilation: the process of framing new experi-
ences through the existent knowledge framework,
without changing it; the structure exists, it is filled
by data;
accomodation: the process of re-framing the
agent’s knowledge framework, usually respond-
ing to contradictions or operational failures of
their knowledge framework; the structure is reor-
ganized.
This theory is aligned with our contribution, as:
rule bases are interpreted as compiled programs
for reactive symbolic processing modules, which
respond to facts (e.g. data fed by sensors), pos-
sibly performing actions (e.g. sending stimuli to
actuators, communicating, etc.);
construction and reconstruction of rule bases pro-
vide the reflective, adaptive processing dimen-
sion, which occurs concurrently to the first one.
The paper is a starting operationalization, there-
fore several research directions remain to be inves-
tigated. First, an evaluation of the application of
our proposal on more complex rule bases, possibly
in comparison with other approaches from the expert
systems literature. Related to that, an in-depth analy-
sis of the proposed algorithms is required, measuring
their complexity, and suggesting possible optimiza-
tions. Second, an investigation of possible theoretical
interactions with the insights coming from belief re-
vision, e.g. (Dubois, 2008), but applied on our defini-
tion of rule revision.
Finally, the paper explores adaptation only as a
problem of maximization of payoffs, studied in deci-
sion theory, game theory and similar top-down per-
spectives. In the future, we plan to analyze it through
the lens of Heiner’s theory of predictable behaviour
(Heiner, 1983); in this way, we expect to be able to
model adaptation as a bottom-up mechanism as well.
REFERENCES
Baral, C. and Gelfond, M. (1994). Logic programming and
knowledge representation. The Journal of Logic Pro-
gramming, 19:73–148.
Boer, A. and Engers, T. (2013). Agile: a problem-based
model of regulatory policy making. Artificial Intelli-
gence and Law, 21(4):399–423.
Dubois, D. (2008). Three scenarios for the revision of epis-
temic states. Journal of Logic and Computation.
Ferm´e, E. (2011). On the Logic of Theory Change: Extend-
ing the AGM Model. PhD thesis, Royal Institute of
Technology.
Goldsmith, J. and Sloan, R. H. (2005). New Horn Revision
Algorithms. Journal of Machine Learning Research,
6:1919–1938.
Heiner, R. (1983). The origin of predictable behavior. The
American economic review, 73(4):560–595.
Horty, J. F. (2011). Rules and Reasons in the Theory of
Precedent. Legal Theory, 17(01):1–33.
Lifschitz, V. (2008). What Is Answer Set Programming?
Proceedings of the AAAI Conference on Artificial In-
telligence.
Mccluskey, E. J. (1956). Minimization of Boolean func-
tions. The Bell System Technical Journal, 35(5):1417–
1444.
Omlin, C. and Giles, C. (1996). Rule revision with recurrent
neural networks. IEEE Transactions on Knowledge
and Data Engineering, 8(1):183–188.
Pearl, J. (1990). System Z: A natural ordering of defaults
with tractable applications to nonmonotonic reason-
ing. Proceedings of the 3rd conference on Theoretical
Aspects of Reasoning about Knowledge (TARK’90),
3:121–135.
Piaget, J., Piercy, M., and Berlyne, D. E. (2001). The
Psychology of Intelligence. Routledge classics. Rout-
ledge.
Shannon, C. E. (1948). A mathematical theory of commu-
nication. The Bell System Technical Journal, 27:379–
423/623–656.
AConstructivistApproachtoRuleBases
547