On the Revision of Action Laws: An Algorithmic Approach

On the Revision of Action Laws: An Algorithmic Approach

Ivan Jos´ Varzinczak
e
Meraka Institute
CSIR, Pretoria, South Africa
ivan.varzinczak@meraka.org.za

Abstract about the behavior of actions in this scenario as a transition
system (Figure 2).
Domain descriptions in reasoning about actions are
¬t, c, h
logical theories and as such they may also evolve.
b b
Given that, knowledge engineers also need revision
tools to incorporate new incoming laws about the b t, c, h
b
t, ¬c, h
dynamic environment. Here we fill this gap by pro- b
viding an algorithmic approach for revision of ac- b
tion laws. We give a well defined semantics that t, ¬c, ¬h
ensures minimal change w.r.t. the original models, Figure 2: A transition system depicting the agent’s knowl-
and show correctness of our algorithms w.r.t. the edge about the dynamics of the coffee machine. b, t, c, and h
semantic constructions. stand for, respectively, buy, token, coffee, and hot.
Now, it may be the case that the agent learns that coffee is
1 Introduction the only hot drink available at the machine, or that even with-
out a token she can still buy, or that all possible executions of
Like any logical theory, action theories in reasoning about
buy should lead to states where ¬token is the case. These are
actions may evolve, and thus need revision methods to ade-
examples of revision with new laws about the dynamics of the
quately accommodate new information about the behavior of
environment under consideration. And here we are interested
actions. Recently, update and contraction-based methods for
in exactly these kinds of theory modification.
action theory change have been defined [Eiter et al., 2005;
The contributions of the present work are as follows:
Herzig et al., 2006; Varzinczak, 2008]. Here we continue
this important though quite new thread of investigation and • What is the semantics of revising an action theory by a
develop a minimal change approach for revising laws of an law? How to get minimal change, i.e., how to keep as
action domain description. much knowledge about other laws as possible?
The motivation is as follows. Consider an agent designed • How to syntactically revise an action theory so that its
to interact with a coffee machine (Figure 1). result corresponds to the intended semantics?
Here we answer these questions.

2 Logical Preliminaries
Our base formalism is multimodal logic Kn [Popkorn, 1994].
2.1 Action Theories in Multimodal K
Let A = {a1 , a2 , . . . , an } be the set of atomic actions of a
domain. To each action a there is associated a modal operator
[a]. P = {p1 , p2 , . . . , pn } denotes the set of propositions, or
atoms. L = {p, ¬p : p ∈ P} is the set of literals. ℓ denotes
a literal and |ℓ| the atom in ℓ.
We use ϕ, ψ, . . . to denote Boolean formulas. F is the set
Figure 1: The coffee deliverer agent.
of all Boolean formulas. A propositional valuation v is a max-
imally consistent set of literals. We denote by v ϕ the fact
Among her beliefs, the agent may know that a coffee is a that v satisfies ϕ. By val(ϕ) we denote the set of all valua-
hot drink, that after buying she gets a coffee, and that with tions satisfying ϕ. |= is the classical consequence relation.
CPL
a token it is possible to buy. We can see the agent’s beliefs Cn(ϕ) denotes all logical consequences of ϕ.

With IP(ϕ) we denote the set of prime implicants [Quine, some (not necessarily fully specified) executability laws for
1952] of ϕ. By π we denote a prime implicant, and atm(π) some actions. These may be incorrect at the starting point,
is the set of atoms occurring in π. Given ℓ and π, ℓ ∈ π but revising wrong executability laws is an approach towards
¯
abbreviates ‘ℓ is a literal of π’. For a given set A, A denotes its solution and one of the aims of this work. With further
its complement. Hence atm(π) denotes P atm(π). information the knowledge engineer will have the chance to
We use Φ, Ψ, . . . to denote complex formulas (possibly change them so that eventually they will correspond to the
with modal operators). a is the dual operator of [a] intuition (cf. Section 3).
( a Φ =def ¬[a]¬Φ). The action theory of our running example will thus be:
A Kn -model is a tuple M = W, R where W is a set of coffee → hot, token → buy ⊤,
valuations, and R maps action constants a to accessibility re- T= ¬coffee → [buy]coffee, ¬token → [buy]⊥,
M
lations Ra ⊆ W × W. Given M , |= p (p is true at world w coffee → [buy]coffee, hot → [buy]hot
w
of model M ) if w
M M
p; |= [a]Φ if |= ′ Φ for every w′ s.t. Figure 2 above shows a Kn -model for the theory T.
w w
′
(w, w ) ∈ Ra ; truth conditions for the other connectives are 2.3 Supra-models
as usual. By M we will denote a set of Kn -models. Sometimes it will be useful to consider models whose possi-
M M
M is a model of Φ (noted |= Φ) if and only if |= Φ for all ble worlds are all the possible states allowed by S :
w
M Definition 1 M = W, R is a big frame of T if and only if:
w ∈ W. M is a model of a set of formulas Σ (noted |= Σ)
M • W = val(S ); and
if and only if |= Φ for every Φ ∈ Σ. Φ is a consequence of
the global axioms Σ in all Kn -models (noted Σ |= Φ) if and
K
• R = a∈A Ra , where
n
M M M M
only if for every M , if |= Σ, then |= Φ. Ra = {(w, w′ ) : ∀.ϕ → [a]ψ ∈ Ea , if |= ϕ then |= ′ ψ}
w w
In Kn we can state laws describing the behavior of actions.
Here we distinguish three types of them. Big frames of T are not unique and not always models of T.
Static Laws A static law is a formula ϕ ∈ F that char- M
Definition 2 M is a supra-model of T iff |= T and M is a
acterizes the possible states of the world. An example is big frame of T.
coffee → hot: if the agent holds a coffee, then she holds a
Figure 3 depicts a supra-model of our example T.
hot drink. The set of static laws of a domain is denoted by S .
¬t, c, h
Effect Laws An effect law for a has the form ϕ → [a]ψ, with b b
ϕ, ψ ∈ F. Effect laws relate an action to its effects, which can
M : b
be conditional. The consequent ψ is the effect that always b t, c, h t, ¬c, h
obtains when a is executed in a state where the antecedent b
b
ϕ holds. An example is token → [buy]hot: whenever the
¬t, ¬c, ¬h t, ¬c, ¬h ¬t, ¬c, h
agent has a token, after buying, she has a hot drink. If ψ is
inconsistent we have a special kind of effect law that we call Figure 3: Supra-model for the coffee machine scenario.
an inexecutability law. For example, ¬token → [buy]⊥ says
that buy cannot be executed if the agent has no token. The set 2.4 Prime Valuations
of effect laws of a domain is denoted by E . An atom p is essential to ϕ if and only if p ∈ atm(ϕ′ ) for
Executability Laws An executability law for a has the form all ϕ′ such that |= ϕ ↔ ϕ′ . For instance, p1 is essential to
CPL
ϕ → a ⊤, with ϕ ∈ F. It stipulates the context in which a ¬p1 ∧ (¬p1 ∨ p2 ). atm!(ϕ) will denote the essential atoms of
is guaranteed to be executable. (In Kn a ⊤ reads “a’s exe- ϕ. (If ϕ is a tautology or a contradiction, then atm!(ϕ) = ∅.)
cution is possible”.) For instance, token → buy ⊤ says that For ϕ ∈ F, ϕ∗ is the set of all ϕ′ ∈ F such that ϕ |= CPL
buying can be executed whenever the agent has a token. The ϕ and atm(ϕ′ ) ⊆ atm!(ϕ). For instance, p1 ∨ p2 ∈ p1 ∗,
′
/
set of executability laws of a domain is denoted by X . as p1 |= p1 ∨ p2 but atm(p1 ∨ p2 ) ⊆ atm!(p1 ). Clearly,
CPL
Given a, Ea (resp. Xa ) will denote the set of only those atm( ϕ∗) = atm!( ϕ∗). Moreover, whenever |= ϕ ↔
effect (resp. executability) laws about a. CPL
ϕ′ , then atm!(ϕ) = atm!(ϕ′ ) and also ϕ∗ = ϕ′ ∗.
Action Theories T = S ∪ E ∪ X is an action theory.
Theorem 1 ([Parikh, 1999]) |= CPL
ϕ ↔ ϕ∗, and
2.2 The Frame, Ramification and Qualification ′ ′ ′
atm(ϕ∗) ⊆ atm(ϕ ) for every ϕ s.t. |= ϕ ↔ ϕ .
CPL
Problems Thus for every ϕ ∈ F there is a unique least set of elemen-
To make the presentation more clear to the reader, we here as- tary atoms such that ϕ may equivalently be expressed using
sume that the agent’s theory contains all frame axioms. How- only atoms from that set. Hence, Cn(ϕ) = Cn(ϕ∗).
ever, all we shall say here can be defined within a formalism Given a valuation v, v′ ⊆ v is a subvaluation. For W a set of
with a solution to the frame and ramification problems like valuations, a subvaluation v′ satisfies ϕ ∈ F modulo W (noted
done by Herzig et al. [2006]. v′ W ϕ) if and only if v ϕ for all v ∈ W such that v′ ⊆ v.
Given the acknowledged difficulty of the qualification !
problem, we do not assume here any a priori solution to it. In- A subvaluation v essentially satisfies ϕ modulo W (v W
ϕ) if
stead, we suppose the knowledge engineer may want to state and only if v W ϕ and {|ℓ| : ℓ ∈ v} ⊆ atm!(ϕ).

Definition 3 Let ϕ ∈ F and W be a set of valuations. A 3.1 Revising a Model by a Static Law
subvaluation v is a prime subvaluation of ϕ (modulo W) if Suppose that our agent discovers that the only hot drink that is
! !
and only if v W ϕ and there is no v′ ⊆ v s.t. v′ W ϕ. served on the machine is coffee. In this case, we might want
to revise her beliefs with the new static law coffee ↔ hot.
A prime subvaluation of a formula ϕ is one of the weakest Considering the model in Figure 3, we see that ¬coffee ∧
states of truth in which ϕ is true. (Notice the similarity with hot is satisfiable. As we do not want this, the first step is to
the syntactical notion of prime implicant [Quine, 1952].) We remove all worlds in which ¬coffee ∧ hot is true. The second
denote all prime subvaluations of ϕ modulo W by base(ϕ, W). step is to guarantee all the remaining worlds satisfy the new
Theorem 2 Let ϕ ∈ F and W be a set of valuations. Then law. This issue has been largely addressed in the literature on
for all w ∈ W, w ϕ if and only if w v∈base(ϕ,W) ℓ∈v ℓ. belief revision and update [G¨ rdenfors, 1988; Winslett, 1988;
a
Katsuno and Mendelzon, 1992; Herzig and Rifi, 1999]. Here
2.5 Closeness Between Models we can achieve that with a semantics similar to that of clas-
When revising a model, we perform a change in its structure. sical revision operators: basically one can change the set of
Because there can be several ways of modifying a model (not possible valuations, by removing or adding worlds.
all of them minimal), we need a notion of distance between In our example, removing the possible worlds {t, ¬c, h}
models to identify those that are closest to the original one. and {¬t, ¬c, h} would do the job (there is no need to add new
As we are going to see in more depth in the sequel, chang- valuations since the new static law is satisfied in at least one
ing a model amounts to modifying its possible worlds or world of the original model).
its accessibility relation. Hence, the distance between two The delicate point in removing worlds is that it may result
Kn -models will depend upon the distance between their sets in the loss of some executability laws: in the example, if there
of worlds and accessibility relations. These here will be were only one arrow leaving some world w and pointing to
based on the symmetric difference between sets, defined as {¬t, ¬c, h}, then removing the latter from the model would
˙
X −Y = (X Y ) ∪ (Y X). make the action under concern no longer executable in w.
Here we claim that this is intuitive: if the state of the world
Definition 4 Let M = W, R . M ′ = W′ , R′ is at least as to which we could move is no longer possible, then we do
close to M as M ′′ = W′′ , R′′ , noted M ′ M M ′′ , iff not have a transition to that state anymore. Hence, if that
• either W−W′ ⊆ W−W′′
˙ ˙ transition was the only one we had, it is natural to lose it.
One could also ask what to do with the accessibility rela-
• or W−W′ = W−W′′ and R−R′ ⊆ R−R′′
˙ ˙ ˙ ˙
tion if new worlds must be added (revision case). We claim
This is an extension of Burger and Heidema’s [2002] relation that it is reckless to blindly add new elements to R. Instead,
to our modal case. Note that other distance notions are also we shall postpone correction of executability laws, if needed.
possible, like e.g. the cardinality of symmetric differences or This approach is debatable, but with the information we have
Hamming distance. (See Section 7 for an explanation on why at hand, it is the safest way of changing static laws.
we have chosen this particular distance notion here.) The semantics for revision of one model by a static law is
as follows:
3 Semantics of Revision
Definition 5 Let M = W, R . M ′ = W′ , R′ ∈ Mϕ iff
⋆
Contrary to contraction, where we want the negation of a law ′ ′
W = (W val(¬ϕ)) ∪ val(ϕ) and R ⊆ R.
to be satisfiable, in revision we want a new law to be valid.
Thus we must eliminate all cases satisfying its negation. M′
Clearly |= ϕ for all M ′ ∈ Mϕ . The minimal models of
⋆

The idea in our semantics is as follows: we initially have a the revision of M by ϕ are those closest to M w.r.t. M :
set of models M in which a given formula Φ is (potentially) ⋆
not valid, i.e., Φ is (possibly) not true in every model in M. Definition 6 rev(M , ϕ) = min{Mϕ , M }.

In the result we want to have only models of Φ. Adding Φ- In the example of Figure 3, rev(M , coffee ↔ hot) is the
models to M is of no help. Moreover, adding models makes singleton {M ′ }, with M ′ as shown in Figure 4.
us lose laws: the resulting theory would be more liberal. ¬t, c, h
One solution amounts to deleting from M those models b
that are not Φ-models. Of course removing only some of
them does not solve the problem, we must delete every such a M′ : b t, c, h
model. By doing that, all resulting models will be models of b
Φ. (This corresponds to theory expansion, when the resulting b
¬t, ¬c, ¬h t, ¬c, ¬h
theory is satisfiable.) However, if M contains no model of
Φ, we will end up with ∅. Consequence: the resulting theory Figure 4: Revising model M in Figure 3 with coffee ↔ hot.
is inconsistent. (This is the main revision problem.) In this
case the solution is to substitute each model M in M by its
⋆
nearest modifications MΦ that makes Φ true. This lets us to 3.2 Revising a Model by an Effect Law
keep as close as possible to the original models we had. Let’s suppose now that our agent eventually discovers that
Before defining revision of sets of models, we present what after buying coffee she does not keep her token. This means
modifications of (individual) models are. that her theory should now be revised by the new effect law

token → [buy]¬token. Looking at model M in Figure 3, this To remove ¬token ∧ [buy]⊥ we have to look at all worlds
amounts to guaranteeing that token ∧ buy token is satisfiable satisfying it and modify M so that they no longer satisfy that
in none of its worlds. To do that, we have to look at all the formula. Given worlds w4 = {¬token, ¬coffee, ¬hot} and
worlds satisfying this formula (if any) and w5 = {¬token, ¬coffee, hot}, we have two options: change
• either make token false in each of these worlds, the interpretation of token in both or add new arrows leaving
these worlds. A question that arises is ‘what choice is more
• or make buy token false in all of them. drastic: change a world or an arrow’? Again, here we claim
If we chose the first option, we will essentially flip the truth that changing the world’s content (the valuation) is more dras-
value of literal token in the respective worlds, which changes tic, as the existence of such a world is foreseen by some static
the set of valuations of the model. If we chose the latter, law and is hence assumed to be as it is, unless we have enough
we will basically remove buy-arrows leading to token-worlds, information supporting the contrary, in which case we explic-
which amounts to changing the accessibility relation. itly change the static laws (see Section 3.1). Thus we shall
In our example, the worlds w1 = {token, coffee, hot}, add a new buy-arrow from each of w4 and w5 .
w2 = {token, ¬coffee, hot} and w3 = {token, ¬coffee, ¬hot} Having agreed on that, the issue now is: which worlds
satisfy token ∧ buy token. Flipping token in all of them to should the new arrows point to? In order to comply with min-
¬token would do the job, but would also have as consequence imal change, the new arrows shall point to worlds that are
the introduction of a new static law: ¬token would now be relevant targets of each of the ¬token-worlds in question.
valid, i.e., the agent never has a token! Do we want this? Definition 9 Let M = W, R , w, w′ ∈ W, and M be a set
We claim that changing action laws should not have as side
of models s.t. M ∈ M. Then w′ is a relevant target world of
effect a change in the static laws. These have a special status, M
and should change only if required (see Section 3.1). Hence w w.r.t. ϕ → a ⊤ for M in M iff |= ϕ and
w
each world satisfying token ∧ buy token has to be changed • If there is M ′ = W′ , R′ ∈ M such that R′ (w) = ∅:
a
so that buy token becomes untrue in it. In the example, we
thus should remove (w1 , w1 ), (w2 , w1 ) and (w3 , w1 ) from R. – for all ℓ ∈ w′ w, there is ψ ′ ∈ F s.t. there is
Mi
The semantics of one model revision for the case of a new v′ ∈ base(ψ ′ , W) s.t. v′ ⊆ w′ , ℓ ∈ v′ , and |= [a]ψ ′
w
effect law is: for every Mi ∈ M
Definition 7 Let M = W, R . M ′ = W′ , R′ ∈ Mϕ→[a]ψ iff:
⋆ – for all ℓ ∈ w ∩ w′ , either there is ψ ′ ∈ F s.t. there is
M
M′
v′ ∈ base(ψ ′ , W) s.t. v′ ⊆ w′ , ℓ ∈ v′ , and |= i [a]ψ ′
w
• W′ = W, R′ ⊆ R, |= ϕ → [a]ψ, and M
for all Mi ∈ M; or there is Mi ∈ M s.t. |= i [a]¬ℓ
w
′ ′ M
• If (w, w ) ∈ R R , then |= ϕ
w • If R′ (w) = ∅ for every M ′ = W′ , R′ ∈ M:
a
The minimal models resulting from the revision of a model – for all ℓ ∈ w′ w, there is Mi = Wi , Ri ∈ M s.t.
M by a new effect law are those closest to M w.r.t. M : there is u, v ∈ Wi s.t. (u, v) ∈ Ria and ℓ ∈ v u
⋆
Definition 8 rev(M , ϕ → [a]ψ) = min{Mϕ→[a]ψ , M }. – for all ℓ ∈ w ∩w′ , there is Mi = Wi , Ri ∈ M s.t.
there is u, v ∈ Wi s.t. (u, v) ∈ Ri a and ℓ ∈ u ∩v, or
Taking M as in Figure 3, rev(M , token → [buy]¬token)
for all Mi = Wi , Ri ∈ M, if (u, v) ∈ Ri a , then
will be the singleton {M ′ } depicted in Figure 5.
¬ℓ ∈ v u
/
¬t, c, h
b b
By rt(w, ϕ → a ⊤, M , M) we denote the set of all relevant
target worlds of w w.r.t. ϕ → a ⊤ for M in M.
′
M : t, c, h t, ¬c, h
In our example, w6 = {¬token, coffee, hot} is the only rel-
b
evant target world here: the two other ¬token-worlds violate
¬t, ¬c, ¬h t, ¬c, ¬h ¬t, ¬c, h the direct effect coffee of action buy, while the three token-
Figure 5: Revising M in Figure 3 with token → [buy]¬token. worlds would make us violate the frame axiom ¬token →
[buy]¬token.
Note that adding effect laws will never require new arrows. The semantics for one model revision by a new executabil-
This is the job of executability-revision. ity law is as follows:
Definition 10 Let M = W, R . M ′ = W′ , R′ ∈ Mϕ→
⋆
iff:
3.3 Revising a Model by an Executability Law a ⊤

Let us now suppose that at some stage it has been decided to M′
• W′ = W, R ⊆ R′ , |= ϕ → a ⊤, and
grant free coffee to everybody. Faced with this information,
we have to revise the agent’s laws to reflect the fact that buy • If (w, w′ ) ∈ R′ R, then w′ ∈ rt(w, ϕ → [a]⊥, M , M)
can also be executed in ¬token-contexts: ¬token → buy ⊤ The minimal models resulting from revising a model M
is a new executability law (and hence we will have buy ⊤ in by a new executability law are those closest to M w.r.t. M :
all new models of the agent’s beliefs). Definition 11 rev(M , ϕ → a ⊤) = ⋆
min{Mϕ→ a ⊤, M }.
Considering model M in Figure 3, we observe that
¬token∧[buy]⊥ is satisfiable. Hence we must throw ¬token∧ In our running example, rev(M , ¬token → buy ⊤) is the
[buy]⊥ away to ensure the new law becomes true. singleton {M ′ }, where M ′ is as shown in Figure 6.

b

b ¬t, c, h b
Algorithm 1 Revision by a Static Law
b b input: T, ϕ
b
M′ : b output: T ⋆ ϕ
t, c, h t, ¬c, h
S ′ := S ⋆ ϕ /* classically revise S */
b ′
E := E /* effect laws remain unchanged */
b
X ′ := ∅ /* executability laws will be ‘recovered’ from old T */
¬t, ¬c, ¬h t, ¬c, ¬h ¬t, ¬c, h
for all π ∈ IP(S ′ ) do
Figure 6: The result of revising model M in Figure 3 by the for all A ⊆ atm(π) do V
new executability law ¬token → buy ⊤. ϕA := pi ∈atm(π) pi ∧ pi ∈atm(π) ¬pi
V
pi ∈A /
pi ∈A
/* by extending π with ϕA we get a valuation */
3.4 Revising Sets of Models if S ′ |= (π ∧ ϕA ) → ⊥ /* context not removed */ then
CPL
if S |= (π ∧ ϕA ) → ⊥ then
CPL
Up until now we have seen what the revision of single models if T |=n (π∧ϕA ) → a ⊤ and S ′ , E ′ , X |=n ¬(π∧ϕA )
K K
means. Now we are ready for a unified definition of revision then
of a set of models M by a new law Φ (cf. Section 5): Xa ′ := {(ϕi ∧ π ∧ ϕA ) → a ⊤ : ϕi → a ⊤ ∈ Xa }
/* preserve executability law in state not removed */
Definition 12 Let M be a set of models and Φ a law. Then else
M E ′ := E ′ ∪ {(π ∧ ϕA ) → [a]⊥}
M⋆ = (M {M :|= Φ}) ∪
Φ rev(M , Φ) T ϕ := S ∪ E ′ ∪ X ′
⋆ ′

M ∈M

Definition 12 comprises both expansion and revision: in the
former, addition of the new law gives a satisfiable theory; in 4.2 Revising a Theory by an Effect Law
the latter a deeper change is needed to get rid of inconsistency.
When revising a theory by a new effect law ϕ → [a]ψ, we
want to eliminate all possible executions of a leading to ¬ψ-
4 Algorithms for Revision of Laws states. To achieve that, we look at all ϕ-contexts and every
We now turn our attention to the syntactical counterpart of time a transition to some ¬ψ-context is not always the case,
revision. Our endeavor here is to perform minimal change i.e., T |= ϕ → a ¬ψ, we can safely force [a]ψ for that con-
Kn
also at the syntactical level. By T ⋆ we denote the result of text. On the other hand, if in such a context there is always
Φ
revising an action theory T with a new law Φ. an execution of a to ¬ψ, then we should strengthen the exe-
cutability laws to make room for the new effect in that context
4.1 Revising a Theory by a Static Law we want to add. Algorithm 2 below does the job.
Looking at the semantics of revision by Boolean formulas,
we see that revising an action theory by a new static law may Algorithm 2 Revision by an Effect Law
conflict with the executability laws: some of them may be lost input: T, ϕ → [a]ψ
and thus have to be changed as well. output: T ⋆ϕ→[a]ψ
The approach here is to preserve as many executability T ′ := T
laws as we can in the old possible states. To do that, we for all π ∈ IP(S ∧ ϕ) do
look at each possible valuation that is common to the new for all A ⊆ atm(π) do V
ϕA := pi ∈atm(π) pi ∧ pi ∈atm(π) ¬pi
V
S and the old one. Every time an executability used to
pi ∈A /
pi ∈A
hold in that state and no inexecutability holds there now,
/* by extending π with ϕA we get a valuation */
we make the action executable in such a context. For those if S |= (π ∧ ϕA ) → ⊥ /* is an allowed context */ then
CPL
contexts not allowed by the old S , we make a inexecutable for all π ′ ∈ IP(S ∧ ¬ψ) do
(cf. Section 3.1). Algorithm 1 deals with that (here S ⋆ ϕ if T ′ |=n (π ∧ ϕA ) → a π ′ /* ¬ψ is achievable */
K
denotes the classical revision of S by ϕ built upon some then
well established method from the literature [Winslett, 1988;
Katsuno and Mendelzon, 1992; Herzig and Rifi, 1999]. The (T ′ X ′ ) ∪ {(ϕi ∧ ¬(π ∧ ϕA )) → a ⊤ :
T ′ := a
choice of a particular operator for classical revision/update ϕi → a ⊤ ∈ X ′ }
a
is not the main matter here, but rather whether it gives us a /* weaken executability laws */
modified set of static laws entailing the new one). T ′ := T ′ ∪ {(π ∧ ϕA ) → [a]ψ} /* safely add the law */
In our example, revising the action theory T with a new if T ′ |=n (π ∧ ϕA ) → [a]⊥ then
K
static law coffee ↔ hot will give us T ′ := T ′ ∪{(ϕi ∧π ∧ϕA ) → a ⊤ : ϕi → a ⊤ ∈ T}
/* preserve other previous transitions */
coffee ↔ hot,
 
  T ϕ→[a]ψ := T ′
⋆

(token ∧ coffee ∧ hot) → buy ⊤,

 

 
⋆
T coffee↔hot = (token ∧ ¬coffee ∧ ¬hot) → buy ⊤,
 ¬coffee → [buy]coffee, ¬token → [buy]⊥, 

 
 In our running example, revision of the action theory T
coffee → [buy]coffee, hot → [buy]hot
 
with the new effect law token → [buy]¬token would give us

T⋆
token→[buy]¬token = Algorithm 3 Revision by an Executability Law
input: T, ϕ → a ⊤
output: T ⋆ a ⊤
coffee → hot, ϕ→
 
  T ′ := T
 (token ∧ ¬(token ∧ coffee ∧ hot)) → buy ⊤,

 

for all π ∈ IP(S ∧ ϕ) do
 


 (token ∧ coffee ∧ hot) → buy ⊤, 

for all A ⊆ atm(π) do V
 
 (token ∧ ¬(token ∧ ¬coffee ∧ hot)) → buy ⊤,
 

ϕA := pi ∈atm(π) pi ∧ pi ∈atm(π) ¬pi

 
 V
(token ∧ ¬coffee ∧ hot) → buy ⊤,
 
pi ∈A /
pi ∈A

 (token ∧ ¬(token ∧ ¬coffee ∧ ¬hot)) → buy ⊤,
  /* by extending π with ϕA we get a valuation */
(token ∧ ¬coffee ∧ ¬hot) → buy ⊤, if S |= (π ∧ ϕA ) → ⊥ /* is an allowed context */ then
 

CPL

 

¬coffee → [buy]coffee, ¬token → [buy]⊥, if T ′ |=n (π ∧ ϕA ) → [a]⊥ then

 


 
 K


 coffee → [buy]coffee, hot → [buy]hot, 

 −
token → [buy]¬token (T ′ E ′ a ) ∪ {(ϕi ∧ ¬(π ∧ ϕA )) → [a]ψi :
 
′−
L i → [a]ψi ∈ E a } ∪
ϕ
T ′ := {(ϕi ∧ π ∧ ϕA ) → [a] ′
π ′ ∈IP(S ) (π ∧ ϕA′ ) :
Regarding the bunch of new executability laws introduced in A′ ⊆atm(π ′ )
the resulting theory, observe that they can be easily simplified −
ϕi → [a]ψi ∈ E ′ a }
to the single one token → buy ⊤.
/* weaken the effect laws */
for all L ⊆ L do
if S |= (π ∧ ϕA ) → ℓ∈L ℓ then
V
4.3 Revising a Theory by an Executability Law CPL
for all ℓ ∈ L do
if T |=n ℓ → [a]⊥ or (T ⊢n ℓ → [a]¬ℓ and
K K
Revision of a theory by a new executability law has as conse- T |=n ℓ → [a]ℓ) then
K
quence a change in the effect laws: all those laws preventing T ′ := T ′ ∪ {(π ∧ ϕA ∧ ℓ) → [a]ℓ}
the execution of a shall be weakened. Moreover, to comply /* preserve non-affected literals */
with minimal change, we must ensure that in all models of T ′ := T ′ ∪ {(π ∧ ϕA ) → a ⊤} /* safely add the law */
′
the resulting theory there will be at most one transition by a T⋆
ϕ→ a ⊤ := T
from those worlds in which T precluded a’s execution.
ϕ,⊥ ϕ,⊥
Let (Ea )1 , . . . , (Ea )n denote minimum subsets (w.r.t. cutability law ¬token → buy ⊤ gives us T ⋆ ¬token→ buy ⊤ =
ϕ,⊥
set inclusion) of Ea such that S , (Ea )i |= ϕ → [a]⊥.
Kn coffee → hot, token → buy ⊤,
 
(According to Herzig and Varzinczak [2007], one can en- 

 ¬coffee → [buy]coffee,



−
sure at least one such a set always exists.) Let Ea =

 

ϕ,⊥ −

 (¬token ∧ ¬(¬token ∧ coffee ∧ hot)∧ 

1≤i≤n (Ea )i . The effect laws in Ea will serve as guide-
 
¬(¬token ∧ ¬coffee ∧ hot)∧

 

 
lines to get rid of [a]⊥ in each ϕ-world allowed by T: they
 
¬(¬token ∧ ¬coffee ∧ ¬hot)) → [buy]⊥,
 
are the laws to be weakened to allow for a ⊤ in ϕ-contexts.  coffee → [buy]coffee, hot → [buy]hot, 
 (¬token ∧ coffee ∧ hot) → [buy]¬token, 

Our algorithm works as follows. To force ϕ → a ⊤ to  

 
 (¬token ∧ ¬coffee ∧ hot) → [buy]¬token, 

be true in all models of the resulting theory, we visit every  


 (¬token ∧ ¬coffee ∧ ¬hot) → [buy]¬token,  
possible ϕ-context allowed by it and make the following op- 
 

¬token → buy ⊤
 
erations to ensure a ⊤ is the case for that context: Given a ϕ-
context, if T does not always preclude a from being executed Again, the resulting theory can be post-processed to give us a
in it, we can safely force a ⊤ without modifying other laws.
much more compact representation of the new laws that have
On the other hand, if a is always inexecutable in that context,
− been added.
then we should weaken the laws in Ea . The first thing we
must do is to preserve all old effects in all other ϕ-worlds. 4.4 Complexity Issues
To achieve that we specialize the above laws to each possible
valuation (maximal conjunction of literals) satisfying ϕ but Algorithms 1–3 terminate. However, they come with a con-
the actual one. Then, in the current ϕ-valuation, we must en- siderable computational cost: the Kn -entailment test with
sure that action a may have any effect, i.e., from this ϕ-world global axioms is known to be EXPTIME-complete. Comput-
we can reach any other possible world. We achieve that by ing all possible contexts allowed by the theory is clearly ex-
−
weakening the consequent of the laws in Ea to the exclusive ponential. Moreover, the computation of IP(.) might result in
disjunction of all possible contexts in T. Finally, to get mini- exponential growth [Marquis, 2000].
mal change, we must ensure that all literals in this ϕ-valuation Given that theory change can be done offline, from the
that are not forced to change are preserved. We do this by stat- knowledge engineer’s perspective what is more important is
ing a conditional frame axiom of the form (ϕk ∧ ℓ) → [a]ℓ, the size of the computed contracted theories. In that matter,
where ϕk is the above-mentioned ϕ-valuation. our results are positive:
Algorithm 3 gives the pseudo-code for that. Theorem 3 Let T be an action theory, and Φ be a law. Then
In our example, revising the action theory T with the exe- the size (number of formulas) of T ⋆ is linear in that of T.
Φ

5 Correctness of the Algorithms feature is important when the qualification problem is consid-
Suppose we have two atoms p1 and p2 , and one action a. Let ered (cf. the Introduction).
T1 = {¬p2 , p1 → [a]p2 , a ⊤}. The only model of T1 is M The work by Eiter et al. [2005; 2006] is similar to ours in
in Figure 7. Revising such a model by p1 ∨ p2 gives us the that they also propose a framework that is oriented to updat-
models Mi′ , 1 ≤ i ≤ 3, in Figure 7. Now, revising T1 by ing action laws. They mainly investigate the case where e.g.
p1 ∨ p2 will give us T1 ⋆1 ∨p2 = {p1 ∧ ¬p2 , p1 → [a]p2 }. The
p a new effect law is added to the description (and then has to
only model of T1 ⋆1 ∨p2 is M1 in Figure 7. This means that the
′ be true in all models of the modified theory).
p
′
semantic revision may produce models (viz. M2 and M3 in ′ In Eiter et al.’s framework, action theories are described
Figure 7) that are not models of the revised theories. in a variant of a narrative-based action description language.
Like in the present work, the semantics is also in terms of
a
transition systems: directed graphs having arrows (action oc-
M : ¬p1 , ¬p2 ′
M1 : p1 , ¬p2 currences) linking nodes (configurations of the world). Con-
trary to us, however, the minimality condition on the outcome
′ ′
of the update is in terms of inclusion of sets of laws, which
M2 : ¬p1 , p2 M3 : p1 , p2
means the approach is more syntax oriented.
Figure 7: Model M of T1 and revision of M by p1 ∨ p2 . In their setting, during an update an action theory T is seen
as composed of two pieces, Tu and Tm , where Tu stands for
The other way round the algorithms may give theories the part of T that is not supposed to change and Tm contains
whose models do not result from revision of models of the the laws that may be modified. In our terms, when revising
initial theory: let T2 = {(p1 ∨ p2 ) → [a]⊥, a ⊤}. Its only by a static law we would have Tm = S ∪ Xa , when revising
model is M (Figure 7). Revising M by p1 ∨ p2 is as above. by an effect law Tm = Ea ∪ Xa , and when revising with exe-
−
But T2 ⋆1 ∨p2 = {p1 ∨ p2 , (p1 ∨ p2 ) → [a]⊥} has a model cutability laws Tm = Ea ∪ Xa . The difference here is that in
p
our approach it is always clear what laws should not change
M ′′ = {{p1 , p2 }, {p1 , ¬p2 }, {¬p1 , p2 }}, ∅ not in Mp⋆ ∨p2 .
1 in a given type of revision, and Tu and Tm do not need to be
All this happens because the possible states are not com- explicitly specified prior to the update.
pletely characterized by the static laws. Fortunately, concen- Their approach and ours can both be described as
trating on supra-models of T, we get the right result. constraint-based update, in that the theory change is carried
Theorem 4 If M = {M : M is a supra-model of T} and out relative to some restrictions (a set of laws that we want to
M′ hold in the result). In our framework, for example, all changes
there is M ′ ∈ M s.t. |= Φ, then M ∈M rev(M , Φ) ⊆ M. in the action laws are relative to the set of static laws S (and
Then, revision of models of T by a law Φ in the semantics that is why we concentrate on supra-models: models of T
produces models of the output of the algorithms T ⋆ :
Φ having val(S ) as worlds). When changing a law, we want to
keep the same set of states. The difference w.r.t. Eiter et al.’s
Theorem 5 If M = {M : M is a supra-model of T} = ∅, approach is that there it is also possible to update a theory
M′
then for every M ′ ∈ M⋆ , |= T ⋆ .
Φ Φ relatively to e.g. executability laws: when expanding T with a
new effect law, one may want to constrain the change so that
Also, models of T ⋆ result from revision of models of T by Φ:
Φ the action under concern is guaranteed to be executable in the
Theorem 6 If M = {M : M is a supra-model of T} = ∅, result.1 As shown in the referred work, this may require the
M′ withdrawal of some static law. Hence, in Eiter et al.’s frame-
then for every M ′ , if |= T ⋆ , then M ′ ∈ M⋆ .
Φ Φ
work, static laws do not have the same status as in ours.
Sticking to supra-models of T is not a big deal. We can use
existent algorithms in the literature [Herzig and Varzinczak, 7 Discussion and Perspectives
2007] to ensure that T is characterized by its supra-models
and that M = ∅. Here we have studied what revising action theories by a law
means, both in the semantics and at the syntactical (algorith-
6 Related Work mic) level. We have defined a semantics based on distances
between models that also captures minimal change w.r.t. the
The problem of action theory change has only recently preservation of effects of actions. With our algorithms and
received attention in the literature, both in action lan- the correctness results we have established the link between
guages [Baral and Lobo, 1997; Eiter et al., 2005] and in the semantics and the syntax for theories with supra-models.
modal logic [Herzig et al., 2006; Varzinczak, 2008]. (Due to page limits, proofs have been omitted here.)
Baral and Lobo [1997] introduce extensions of action lan- For the sake of presentation, here we have abstracted from
guages that allow for some causal laws to be stated as defea- the frame and ramification problems. However our definitions
sible. Their work is similar to ours in that they also allow for could have been stated in a formalism with a suitable solution
weakening of laws: in their setting, effect propositions can be to them, like e.g. Castilho et al.’s approach [1999]. With re-
replaced by what they call defeasible (weakened versions of) gards to the qualification problem, this is not ignored here:
effect propositions. Our approach is different from theirs in
the way executability laws are dealt with. Here executability 1
We can emulate that in our approach with two modifications
laws are explicit and we are also able to change them. This of T: first adding the effect law and then an executability law.

revising wrong executability laws is an approach towards its [Burger and Heidema, 2002] I.C. Burger and J. Heidema.
solution. Indeed, given the difficulty of stating all sufficient Merging inference and conjecture by information. Syn-
conditions for executability of an action, the knowledge en- these, 131(2):223–258, 2002.
gineer writes down some of them and lets the theory ‘evolve’ [Castilho et al., 1999] M. Castilho, O. Gasquet, and
via subsequent revisions. A. Herzig. Formalizing action and change in modal
The reason why we have chosen such a simple notion of logic I: the frame problem. J. of Logic and Computation,
distance between models is that with other distances one may 9(5):701–735, 1999.
not always get the intended result. This is better illustrated [Dalal, 1988] M. Dalal. Investigations into a theory of
with the contraction counterpart of our operators [Varz-
knowledge base revision: preliminary report. In Proc.
inczak, 2008]. Suppose one wants to remove an executability
AAAI, pages 475–479, 1988.
law ϕ → a ⊤. Then we do that by removing a-arrows from
ϕ-worlds. Suppose we have a model with two ϕ-worlds, w1 [Eiter et al., 2005] T. Eiter, E. Erdem, M. Fink, and J. Senko.
with one leaving a-arrow and w2 with two a-arrows. Then Updating action domain descriptions. In Proc. IJCAI,
with e.g. Dalal’s distance [1988], the associated contraction pages 418–423, 2005.
operator would always exclude the resulting model in which [Eiter et al., 2006] T. Eiter, E. Erdem, M. Fink, and J. Senko.
w2 loses its two arrows, simply because deleting 1 arrow Resolving conflicts in action descriptions. In Proc. ECAI,
is Dalal-better than deleting 2. This problem doesn’t hap- pages 367–371, 2006.
pen with our distance, which gives us a version of maxi-
[G¨ rdenfors, 1988] P. G¨ rdenfors. Knowledge in Flux: Mod-
a a
choice [Hansson, 1999].
eling the Dynamics of Epistemic States. MIT Press, 1988.
One criticism to the approach here developed concerns the
precedence of static laws in the revision process, which could [Hansson, 1999] S. Hansson. A Textbook of Belief Dynam-
make the revision operators to be interpreted as incoherent. ics: Theory Change and Database Updating. Kluwer,
As agreed in the literature, however, given that static laws are 1999.
much easier to state, they are more likely to be correct, and [Herzig and Rifi, 1999] A. Herzig and O. Rifi. Propositional
then it makes sense to give them precedence. Supporting this belief base update and minimal change. Artificial Intelli-
is the fact that most of the attention in the reasoning about gence, 115(1):107–138, 1999.
actions area has been paid to effect laws and executability [Herzig and Varzinczak, 2007] A. Herzig and I. Varzinczak.
laws, which are much more difficult to completely specify. Metatheory of actions: beyond consistency. Artificial In-
Our approach is in line with that. telligence, 171:951–984, 2007.
Our next step is to analyze the behavior of our opera- [Herzig et al., 2006] A. Herzig, L. Perrussel, and I. Varz-
tors w.r.t. AGM-like postulates [Alchourr´ n et al., 1985]
o
inczak. Elaborating domain descriptions. In Proc. ECAI,
for modal theories and the relationship between our revision
pages 397–401, 2006.
method and contraction. What is known is that Levi iden-
tity [Levi, 1977], T ⋆ = T − ∪ {Φ}, in general does not hold
Φ ¬Φ
[Katsuno and Mendelzon, 1992] H. Katsuno and A. Mendel-
for action laws. The reason is that up to now there is no con- zon. On the difference between updating a knowledge base
traction operator for ¬Φ where Φ is an action law. Indeed this and revising it. In Belief revision, pages 183–203. Cam-
is the general contraction problem for action theories: con- bridge, 1992.
traction of a theory T by a general formula (like ¬Φ above) is [Levi, 1977] I. Levi. Subjunctives, dispositions and chances.
still an open problem in the area. The definition of a general Synthese, 34:423–455, 1977.
method will mostly benefit from the semantic modifications
[Marquis, 2000] P. Marquis. Consequence finding algo-
we studied here (addition/removal of arrows and worlds).
rithms. In Alg. for Defensible and Uncertain Reasoning,
Given the relationship between modal logics and descrip- pages 41–145. 2000.
tion logics, a revision method for DL TBoxes [Baader et al.,
2003] would also benefit from the constructions that we have [Parikh, 1999] R. Parikh. Beliefs, belief revision, and split-
defined here. ting languages. In Logic, Language and Computation,
pages 266–278, 1999.
References [Popkorn, 1994] S. Popkorn. First Steps in Modal Logic.
[Alchourr´ n et al., 1985] C. Alchourr´ n, P. G¨ rdenfors, and
o o a Cambridge University Press, 1994.
D. Makinson. On the logic of theory change: Partial meet [Quine, 1952] W. V. O. Quine. The problem of simplifying
contraction and revision functions. J. of Symbolic Logic, truth functions. American Mathematical Monthly, 59:521–
50:510–530, 1985. 531, 1952.
[Baader et al., 2003] F. Baader, D. Calvanese, D. McGuin- [Varzinczak, 2008] I. Varzinczak. Action theory contraction
ness, D. Nardi, and P. Patel-Schneider, editors. Descrip- and minimal change. In Proc. KR, pages 651–661, 2008.
tion Logic Handbook. Cambridge University Press, 2003. [Winslett, 1988] M.-A. Winslett. Reasoning about action us-
[Baral and Lobo, 1997] C. Baral and J. Lobo. Defeasible ing a possible models approach. In Proc. AAAI, pages 89–
specifications in action theories. In Proc. IJCAI, pages 93, 1988.
1441–1446, 1997.

On the Revision of Action Laws: An Algorithmic Approach

More Related Content

Viewers also liked (10)

Similar to On the Revision of Action Laws: An Algorithmic Approach (15)

More from Ivan Varzinczak (11)

Recently uploaded (20)

On the Revision of Action Laws: An Algorithmic Approach