Regression in Modal Logic

Robert Demolombe* — Andreas Herzig** — Ivan Varzinczak**

* ONERA Toulouse
2 Avenue Edouard Belin, B.P. 4025
31055 Toulouse Cedex France
[email protected]
** Institut de Recherche en Informatique de Toulouse
118, route de Narbonne
31062 Toulouse Cedex 4 France
{herzig,ivan}@irit.fr

ABSTRACT. In this work we propose an encoding of Reiter’s Situation Calculus solution to the
frame problem into the framework of a simple multimodal logic of actions. In particular we
present the modal counterpart of the regression technique. This gives us a theorem proving
method for a relevant fragment of our modal logic.
KEYWORDS: reasoning about actions, regression, modal logic, dependence.

1. Introduction

In the reasoning about actions field most approaches use the Situation Calculus
formalism [MCC 69]. Among those, Reiter’s [REI 91] has turned out to be most fruit-
ful. His basic formalism is restricted to deterministic actions without ramifications.
In order to solve the frame problem he makes use of so-called successor state axioms
(SSAs). The latter enable regression [REI 91], which has interesting computational
properties.
The Situation Calculus is a dialect of predicate logic, having situations and actions
as objects, and where actions are viewed as mappings on the set of situations. At first
glance this is very close to possible worlds semantics for Deterministic PDL [HAR 84].
But the precise relation between Reiter’s approach and dynamic logic is not as obvious
as that. One of the reasons why his formalism cannot be translated straightforwardly
into modal logics of action such as PDL is that the Situation Calculus allows quanti-
fying over actions. Worse, such quantifications are central to Reiter’s approach.

Journal of Applied Non-Classical Logics.Volume 13 – n◦ 2/2003, pages 165 to 185

166 Journal of Applied Non-Classical Logics. Volume 13 – n◦ 2/2003

In [DEM 03] there has been presented a technique to translate Reiter’s approach
into dynamic logic. In this paper we present a different approach. We solve the prob-
lem using an extension of dynamic logic that has been introduced in [CAS 99]. There,
dynamic logic is combined with a causal notion based on a dependence relation, re-
sulting in a family of logics LAP ; . LAP ; is a simple yet powerful account to
the frame and ramification problems, with the advantage of having a decision proce-
dure in terms of tableau systems (while the Situation Calculus contains second-order
axioms and is a priori not even semi-decidable). We propose an encoding of Reiter’s
approach into the formalism of LAP ; . Having such a result provides some degree
of optimization in doing inference tasks for some classes of problems in the area.
This work is organized in the following way: in Section 2 we present a slightly
modified version of PDL, which will serve as the basis for developing the central
ideas of this paper. Section 3 is devoted to introduce the basic hypotheses concerning
the knowledge we have about actions. In Section 4 we present Reiter’s solution to the
frame problem in the logical basis of Section 2 and in Section 5 we summarize Reiter’s
regression technique. We then revisit De Giacomo and Lenzerini’s account for encod-
ing domain descriptions into a variant of dynamic logic that avoids quantification over
actions (Section 6). In Section 7 we present our modal logic of actions LAP ; . In
Section 8 we show how we can do regression in LAP ; . Finally we sketch possible
extensions to this work (Section 9) and then give some concluding remarks.
A preliminary version of this paper has been presented at the workshop ‘Methods
for Modalities’ (M4M-3). Thanks are due to Jérôme Lang for comments and discus-
sions.

2. Deterministic PDL with quantification and equality

In this section we introduce a slightly extended version of deterministic PDL con-

taining quantification over actions and the equality predicate.
We use P1 , P2 , . . . for propositional constants, and P, Q, . . . as metavariables for
propositional constants. A1 , A2 , . . . denote action constants and a, b, . . . action vari-
ables. We will use A, B, . . . as metavariables ranging over action constants and vari-
ables. PRP is the set of all propositional constants, and ACT is the set of all action
constants. Examples of propositional constants are Loaded (“the gun is loaded”) and
Alive (“the agent is alive”). Examples of action constants are shoot (“shooting the
agent”) and strangle (“strangling the agent”). L1 , L2 , . . . denote literals. If L = ¬P
then we shall identify ¬L with P . φ, ψ, . . . denote formulas that are constructed in the
usual way from PRP using the classical propositional operators. We shall also call
them classical formulas. Hence they do not contain modal operators, quantifiers or
the equality predicate that is to be introduced below.
We will use modal operators [A] (resp. [a]) for each action constant A ∈ ACT
(resp. action variable a). Φ, Ψ will denote complex formulas possibly involving modal
operators, quantification, and equality between actions. [A]Φ is read “after executing
 Regression in Modal Logic 167

A, Φ”. We also use the dual hAi of [A]. The formula hAi> can be read as “A is
executable”.
The nonstandard feature of our logic is that we allow for quantification over ac-
tions, and for equality between actions. Hence, in this version of dynamic logic we
allow for formulas of the form ∀aΦ, with Φ a complex formula as defined above. In
the Yale shooting scenario (YSS) [HAN 86], one can e.g. write

∀a(Alive ∧ ¬[a]Alive → (a = shoot ∧ HasGun ∧ Loaded)).

This is an explanation closure axiom [SCH 90] expressing that the only way to make
Alive false is by the shooting action.
A model is a triple M = hW, R, Ii where W is a set of Kripke possible worlds, R
is a set of binary relations on W , and I is an interpretation function mapping propo-
sitional constants to subsets of W , and action constants and variables to elements
of R. We will sometimes write w 0 ∈ (I(A))(w) instead of wI(A)w 0 , and similarly
for variables a.
We say that the interpretation I agrees with I 0 except possibly on a if and only if
– I(P ) = I 0 (P ) for every propositional constant P ;
– I(A) = I 0 (A) for every action constant A;
– I(b) = I 0 (b) for every action variable b different from a.
For a given model M = hW, R, Ii, w |=M ∀aΦ if for every I 0 such that I
agrees with I 0 except possibly on a, w |=hW,R,I 0 i Φ. w |=M [A]Φ if for every
w0 ∈ (I(A))(w), w 0 |=M Φ. w |=M [a]Φ if for every w 0 ∈ (I(a))(w), w 0 |=M Φ.
We say that a formula Ψ is a consequence of the set of global axioms {Φ 1 , . . . , Φn }
in the class of models M (noted {Φ1 , . . . , Φn } |=M Ψ) if and only if for all M ∈ M,
if |=M Φi for every Φi , then |=M Ψ.
We will use K to denote the class of all possible models. DK = {hW, R, Ii ∈
K : R is a partial function} is the class of models where actions are deterministic, i.e.,
(I(A))(w) is either a singleton or empty. Thus, for all action constants A and all
formulas Φ
|=DK hAiΦ → [A]Φ (1)

If all actions are deterministic, then every formula without quantification can be
brought into a normal form where there are neither conjunctions nor disjunctions in the
scope of modal operators. Apart from classical equivalences, this uses the following
ones from the left to the right:

|=DK [A](Φ ∧ Ψ) ↔ ([A]Φ ∧ [A]Ψ) (2)

|=DK [A](Φ ∨ Ψ) ↔ ([A]Φ ∨ [A]Ψ) (3)

168 Journal of Applied Non-Classical Logics. Volume 13 – n◦ 2/2003

3. Describing actions

Reiter (and more generally the reasoning about actions community) focuses on
deductions from a theory describing a given set of actions in terms of preconditions
and effects. In dynamic logic such an action theory corresponds to a set of global
axioms in Fitting’s sense [FIT 83]. We have for example,
{[load]Loaded, Loaded → [shoot]¬Alive} |=K [load][shoot]¬Alive

In the Situation Calculus, the same result is obtained by quantifying over situa-
tions. For our running example we have ∀sLoaded(do(load, s)) and ∀s(Loaded(s) →
¬Alive(do(shoot, s))), where s is a variable of sort situation, do is a function symbol,
load, shoot are constants of sort action, and Loaded, Alive predicate symbols.
In describing an action theory it is more or less explicitly supposed that the follow-
ing pieces of information are given. About these items some assumptions of complete
information are made.

3.1. Action preconditions

For each action constant A there is a classical formula Poss(A) describing the ac-
tion precondition of A, i.e. the condition under which A can be executed. For example
Poss(shoot) = HasGun, and Poss(strangle) = >.
It is supposed that the action preconditions are complete: A is executable if and
only if Poss(A) is true.
In terms of dynamic logic, completeness of action preconditions means that for
every A ∈ ACT we have a global axiom Poss(A) ↔ ¬[A]⊥.

3.2. Set of possible causes

For each propositional constant P there are two finite sets of action constants
Cause + (P ) and Cause − (P ) describing the positive and negative causes of P . (Note
that ACT may be infinite.) Cause + (P ) contains the actions in ACT which in some
circumstances might cause P to become true, while Cause − (P ) contains those ac-
tions that may cause P false. For example Cause + (Alive) = ∅ (no action makes an
agent alive), Cause − (Alive) = {shoot, strangle}, and Cause − (Loaded) = {shoot}.1
It is also supposed that Cause + (P ) and Cause − (P ) are small, in the sense that
Cause + (P ) and Cause − (P ) are much smaller than ACT .
Moreover, we suppose that these two sets are complete: whenever A 6∈
Cause + (P ) then the execution of A can never make P true. In terms of dynamic

1. In Reiter’s presentation these functions retrieved from his functions γ + and γ − .

 Regression in Modal Logic 169

logic, causal completeness means that we have a global axiom ¬P → [A]¬P in

that case. Similarly, for every B such that B 6∈ Cause − (P ) we have a global ax-
iom P → [B]P . Axioms of that form are called frame axioms. In our example, as
strangle 6∈ Cause − (Loaded), we have Loaded → [strangle]Loaded.
The next piece of information specifies the causal relation in more detail.

3.3. Effect preconditions

For all propositional constant P ∈ PRP and every action constant A ∈

Cause + (P ) there is a classical formula Cond + (A, P ) describing the positive ef-
fect precondition of action A. Similarly, for every A ∈ Cause − (P ) there
is a Cond − (A, P ) describing its negative effect precondition. For example
Cond − (strangle, Alive) = >, and Cond − (shoot, Alive) = Loaded.2
It is supposed that the effect preconditions are complete: in situations where the
formula Cond + (A, P ) does not hold the execution of A can never make P true. Sym-
metrically, when Cond − (A, P ) does not hold then the execution of A can never make
P false.
In terms of dynamic logic, to every effect precondition Cond + (A, P ) one can
associate a global axiom Cond + (A, P ) → [A]P , and to every effect precondition
Cond − (A, P ) one can associate a global axiom Cond − (A, P ) → [A]¬P . As an
example, consider the formula Loaded → [shoot]¬Alive.
Completeness of effect preconditions means that we moreover have a global axiom
(¬Cond + (A, P ) ∧ ¬P ) → [A]¬P for every A ∈ Cause + (P ). Symmetrically, for
every B such that B ∈ Cause − (P ) we have a global axiom (¬Cond − (B, P )∧P ) →
[B]P . For example we have (¬Loaded ∧ Alive) → [shoot]Alive.

3.4. Comments

The last two completeness assumptions of Sections 3.2 and 3.3 express in modal
logic what Reiter calls “explanation closure” and “Clark completion”.
Most importantly, the three pieces of information together with the completeness
assumptions make that the possible world resulting from the execution of action A in
a possible world w is completely determined: for every model M and world w of M ,
if w 6|=M Poss(A) then (I(A))(w) = ∅. Else the truth value of every P in every w 0
accessible from w via I(A) is as follows. Suppose w.l.o.g. that w |= M P . Then:
– if A 6∈ Cause − (P ) then w0 |=M P ;
– if A ∈ Cause − (P ) and w 6|=M Cond − (A, P ) then w0 |=M P ;
– if A ∈ Cause − (P ) and w |=M Cond − (A, P ) then w0 6|=M P .

2. These functions correspond to Reiter’s γ + and γ − .

170 Journal of Applied Non-Classical Logics. Volume 13 – n◦ 2/2003

As all truth values are thus determined, it follows that the set of worlds accessible
via I(A) is either empty, or it can be considered to be a singleton. This fits with the
assumption that all actions are deterministic.
As we have noted, the action preconditions and effect preconditions appear ex-
plicitly in Reiter’s formalization, while the sets of possible causes Cause + (P ) and
Cause − (P ) only appear implicitly there.
Note that in Reiter’s Situation Calculus it is supposed that actions always lead to
some state: even in states where the agent has no gun in his hands, the state resulting
from the execution of shoot exists. The technical reason is that just as every function
in predicate logic, his successor function do is total. This means that the logic of
each action operator [A] should be KD. We have nevertheless decided to follow the
dynamic logic tradition and suppose that the set of worlds accessible via some action
A might be empty. Therefore the logic of each [A] is just K.
In fact, inexecutability of the action shoot is expressed in Situation Calculus by
stating Poss(shoot) ↔ HasGun, where Poss(shoot) is a particular propositional con-
stant. In our formulation, Poss is a function associating a classical formula to every
action A. Poss(A) can be seen as an abbreviation, such as Poss(shoot) = HasGun.
Given a domain description in Reiter’s style, we obtain a description in our style if we
(1) define our Poss-function from Reiter’s preconditions Poss(A) ↔ φ, and (2) re-
place Reiter’s constants Poss(A) by our hAi>. The other way round, our version
can be translated to Reiter’s by (1) defining his preconditions Poss(A) ↔ φ from
our Poss-function, and (2) recursively replacing [A]φ by Poss(A) → [A]φ. Ob-
serve that the latter is nothing but the well-known translation from modal logic K to
KD [OHL 91, OHL 93].
All this sounds as if action theories could be described in deterministic PDL in a
satisfactory manner, but we have not solved the frame problem yet: as by hypothesis
Cause + (P ) and Cause − (P ) are small, it follows that the size of the set of frame
axioms that we have to state is close to card (PRP ) × card (ACT ). This is usually
considered to be too big, and a central element in the research program of the reason-
ing about actions community was to design mechanisms allowing to infer such frame
axioms without stating them explicitly.
There was a 20-years-long debate about semantics and theorem proving methods
allowing such inferences. Reiter’s proposal seems to have closed the debate at least
in what concerns deterministic actions without side-effects (also called ramifications).
This is going to be presented in the sequel.

4. Reiter’s solution to the frame problem

Based on a particular class of models, Reiter proposes to incorporate the basic

ingredients of action theories that we have presented in the preceding section into
 Regression in Modal Logic 171

successor state axioms which given a state and an action completely determine the
next state.

4.1. Reiter models

Reiter requires that names are unique and that models are trees. Thus,Sgiven a
model M = hW, R, Ii, we say that M is a Reiter model if and only if hW, r∈R ri
is a tree, and if I(Ai ) = I(Aj ), then i = j. RT R will denote the class of all Reiter
models.

4.2. Successor state axioms

Suppose that all the Poss(A), Cause + (P ), Cause − (P ), Cond + (A, P ) and
Cond − (A, P ) are given, and that the completeness assumptions are made. We then
can associate with that an action theory R from which the relevant frame axioms will
follow. In dynamic logic R is made of the following axioms:
– for every A ∈ ACT , there is an executability axiom Poss(A) ↔ ¬[A]⊥;
– for every P ∈ PRP , if Cause + (P ) = {A1 , . . . , An } and Cause − (P ) =
{B1 , . . . , Bm } then there is a successor state axiom

∀a([a]P ↔
(¬Poss(a) ∨
(a = A1 ∧ Cond + (A1 , P )) ∨ . . . ∨ (a = An ∧ Cond + (An , P )) ∨
(P ∧ ¬(a = B1 ∧ Cond − (B1 , P )) ∧ . . . ∧ ¬(a = Bm ∧ Cond − (Bm , P )))))

Note that the successor state axiom is well defined because we have supposed that
Cause + (A) and Cause − (A) are finite.
For the cases where n = 0 or m = 0, conjunction of the elements of an empty
set is identified with >, and disjunction with ⊥. The latter can be illustrated with our
running example, where Cause + (Alive) = ∅. The successor state axiom for Alive is:

∀a([a]Alive ↔
(¬Poss(a) ∨ ⊥ ∨ (Alive ∧ ¬(a = shoot ∧ Loaded) ∧ ¬(a = strangle ∧ >))))

We abbreviate Reg(a, P ) the right hand side of the equivalence. The successor
state axiom for P therefore has the form ∀a([a]P ↔ Reg(a, P )).
172 Journal of Applied Non-Classical Logics. Volume 13 – n◦ 2/2003

Successor state axioms can be equivalently stated for negative literals as:

∀a([a]¬P ↔
(¬Poss(a) ∨ (a = B1 ∧ Cond − (B1 , P )) ∨ . . . ∨ (a = Bm ∧ Cond − (Bm , P )) ∨
(¬P ∧ ¬(a = A1 ∧ Cond + (A1 , P )) ∧ . . . ∧ ¬(a = An ∧ Cond + (An , P )))))

We abbreviate Reg(a, ¬P ) the right hand side of this equivalence. For example the
successor state axiom for ¬Alive is:

∀a([a]¬Alive ↔
(¬Poss(a) ∨ (a = shoot ∧ Loaded) ∨ (a = strangle ∧ >) ∨ (¬Alive ∧ ¬⊥)))

4.3. Comments

Reiter’s original axiom [REI 91] is slightly different from ours:

∀a(Poss(a) → ([a]P ↔
((a = A1 ∧ Cond + (A1 , P )) ∨ . . . ∨ (a = An ∧ Cond + (An , P )) ∨
(P ∧ ¬(a = B1 ∧ Cond − (B1 , P )) ∧ . . . ∧ ¬(a = Bm ∧ Cond − (Bm , P )))))

Our version can be proved to be equivalent to his.

In his book [REI 01] Reiter excluded the precondition Poss(a) from the right hand
side Reg(a, P ) of the SSA, and just writes

∀a([a]P ↔
((a = A1 ∧ Cond + (A1 , P )) ∨ . . . ∨ (a = An ∧ Cond + (An , P )) ∨
(P ∧ ¬(a = B1 ∧ Cond − (B1 , P )) ∧ . . . ∧ ¬(a = Bm ∧ Cond − (Bm , P ))))

Therefore we would have e.g. [shoot]¬Alive ↔ (Loaded∨(¬Alive∧¬⊥)), from which

it follows by classical principles that ¬HasGun ∧ Alive ∧ [shoot]¬Alive → Loaded.
This means that such SSAs do not take into account inexecutability: this issue must
be managed “by hand” by introducing Poss(shoot) atoms in the right places when
proving consequences of SSAs in their recent version.
Finally, we note that Reiter’s presentation also contains precondition axioms of the
form Poss(A) ↔ φ. This is not needed here because we view Poss(A) as a function
returning a classical formula φ, which is directly integrated into our successor state
axiom (cf. also our comments in section 3.4).

5. Reiter’s regression

Successor state axioms are crucial when it comes to the reasoning aspect of the
frame problem, to which we turn now.
 Regression in Modal Logic 173

Given a Reiter’s style action theory R, what can be deduced from it? Suppose
Φ is a complex formula without quantification, action variables, and equality, such
as HasGun → [load][shoot]¬Alive. In order to decide whether R |= RT R Φ, Reiter
proposes to rewrite Φ using the successor state axioms from the left to the right. This
is what he calls regression, and it consists in syntactical substitutions whose iteration
reduces a given formula with action symbols into another one with just propositional
constants.
At each regression step we have to put formulas in normal form such that there
are neither conjunctions nor disjunctions in the scope of modal operators (using the
hypothesis that all actions are deterministic). Hence the innermost modal opera-
tors have just literals in their scope. For the above example, Φ gets ¬HasGun ∨
[load][shoot]¬Alive.
A LGORITHM 1 (R EITER ’ S REGRESSION ). —

input:
– a formula without variables Φ.
– Poss(A), Cause + (P ), Cause − (P ), Cond + (A, P ) and Cond − (A, P ).
output: a classical formula REG(Φ).

begin
while Φ is not classical
put Φ in normal form
choose a subformula [A]L, where L is either P or ¬P , for P ∈ PRP
if L = P then replace [A]P by Reg(A, P )
else replace [A]¬P by Reg(A, ¬P ).
end

Notice that the action variable a of the successor state axiom is instantiated by the
constant A.
In our example, the regression of the subformula [shoot]¬Alive is
¬HasGun ∨ (shoot = shoot ∧ Loaded)∨
(shoot = strangle ∧ >) ∨ (¬Alive ∧ ¬⊥)
This can be simplified to ¬HasGun ∨ Loaded ∨ ¬Alive. Hence the result of the regres-
sion of Φ is ¬HasGun ∨ [load](¬HasGun ∨ Loaded ∨ ¬Alive).
Each rewriting step thus eliminates a modal operator, and iterated application re-
sults in a formula without modal operators. If we iterate regression in our example,
we first put the formula
¬HasGun ∨ [load](¬HasGun ∨ Loaded ∨ ¬Alive)
174 Journal of Applied Non-Classical Logics. Volume 13 – n◦ 2/2003

into normal form, obtaining

¬HasGun ∨ [load]¬HasGun ∨ [load]Loaded ∨ [load]¬Alive.
The regression of subformula [load]¬HasGun is equivalent to ¬HasGun, that of sub-
formula [load]Loaded to >, and that of [load]¬Alive to ¬Alive. We therefore obtain
¬HasGun ∨ ¬HasGun ∨ > ∨ ¬Alive,
which is valid in classical propositional logic. Thus the original formula HasGun →
[load][shoot]¬Alive can be deduced from Reiter’s action theory R.
As regression is proved to be sound [REI 01, Theorem 4.5.2], checking validity
of the original formula amounts to checking satisfiability of the regressed one in the
initial state of the world:
T HEOREM 2. — R |=RT R Φ ↔ REG(Φ).
C OROLLARY 3. — R |=RT R Φ if and only if REG(Φ) is valid in Classical Proposi-
tional Logic.
The rest of the paper explores whether regression can be performed in a simpler
framework, in particular without quantifying over actions.

6. De Giacomo and Lenzerini’s encoding into PDL

Reiter’s Situation Calculus based solution has raised the natural question of at what
extent it could be possible to do the same in dynamic logic. Given the expressiveness
limitations of the latter (originally it did not allow for quantification over actions),
many researchers [ZHA 01] have turned to other ways of facing the problems in the
area. There has been others [GIA 95], however, who have tried on the first steps in
that direction.
De Giacomo and Lenzerini have expressed Reiter’s solution in a slightly modified
version of PDL. This is what we take up in this section.
Here we simplify their account a bit and suppose that the set of atomic actions
is the finite ACT = {A1 , A2 , . . . , An }. Then their approach can be said to have the
following ingredients (α, β, . . . denote complex actions):
– Nondeterministic choice α ∪ β;
– Converse α− ;
– A particular nondeterministic atomic action any that can be thought of as the
nondeterministic composition of all atomic actions of ACT: any = A1 ∪A2 ∪. . .∪An ;
– Complement ¬α w.r.t. any, where α = B1 ∪ . . . ∪ Bm for some B1 , . . . , Bm ∈
ACT.
Moreover it is supposed that the past is deterministic, as expressed by the logical
axiom ¬[any − ]¬Φ → [any− ]Φ.
 Regression in Modal Logic 175

Consider our example theory. Its representation in De Giacomo and Lenzerini’s

framework is:

[any](¬Alive → hany− i¬Alive ∨ hshoot− iLoaded ∨ hstrangle− i>)

[any](Alive → hany− iAlive)

Just as for PDL, reasoning in De Giacomo and Lenzerini’s logical framework is

EXPTIME-complete. While their encoding certainly preserves the spirit of Reiter’s
successor state axioms, they did not give the counterpart of Reiter’s regression, and
hence did not investigate whether reasoning for syntactically restricted theories is
“cheaper” than EXPTIME.
In the next section we show how this can be simulated without quantification in a
simple modal logic of actions augmented by a dependence relation.

7. Solving the frame problem without quantification

7.1. Adding dependence information to PDL

In [CAS 99] we have augmented a very simple version of PDL (basically multi-
modal K) with metalogical causal information represented by a dependence relation
; between actions and literals. A;L means “action A may cause literal L”. The
nonexistence of such a A;L in ; (noted A6;L) means that “L will never get true
due to A”.3
A;P is just another way of writing down that A ∈ Cause + (P ), and A;¬P that
A ∈ Cause − (P ).
Suppose ; is given. Semantically, if I(A) is the accessibility relation associated
to action A, the relation ; constrains possible worlds models in the following way:
– if A6;P and w 0 ∈ (I(A))(w) and w 6∈ I(P ) then w 0 6∈ I(P );
– if A6;¬P and w 0 ∈ (I(A))(w) and w ∈ I(P ) then w 0 ∈ I(P ).
The resulting class of models is called LAP ; . We note DLAP ; the class
of LAP ; -models whose accessibility relations are deterministic. It has been shown
in [CAS 99] that the validities of LAP ; are completely axiomatized by the following
set of logical axioms:
1) Some axiomatization of classical logic;
2) [A]Φ ∧ [A](Φ → Ψ) → [A]Ψ;
3) ¬L → [A]¬L if A6;L.

3. In [CAS 99] the language moreover contained an S4 modal operator 2 which implies all
action operators [A]. Laws were prefixed with 2, e.g. 2(Loaded → [shoot]¬Alive). Here we
shall achieve the same thing by viewing action laws as global axioms.
176 Journal of Applied Non-Classical Logics. Volume 13 – n◦ 2/2003

plus the Modus Ponens and the necessitation rule.

It has moreover been shown that LAP ; is decidable and EXPTIME -complete,
and a tableau theorem proving method has been given.

7.2. Solving the frame problem in LAP ;

Suppose all the ingredients Poss(A), Cause + (P ), Cause − (P ), Cond + (A, P ),

Cond − (A, P ) are given, and let us make the completeness assumptions as introduced
in Section 3. We define a dependence relation and a set of global axioms S as follows:
– for every Ai ∈ Cause + (P ) we put Ai ;P , and for every Bj ∈ Cause − (P ) we
put Bj ;¬P ;
– for every A ∈ ACT , add an executability axiom to S:

Poss(A) ↔ ¬[A]⊥ (4)

– for every P ∈ PRP and every Ai ∈ Cause + (P ) add two effect axioms to S:

Cond + (Ai , P ) → [Ai ]P (5)

(¬Cond + (Ai , P ) ∧ ¬P ) → [Ai ]¬P (6)

– for every P ∈ PRP and every Bj ∈ Cause − (P ) add two effect axioms to S:

Cond − (Bj , P ) → [Bj ]¬P (7)

(¬Cond − (Bj , P ) ∧ P ) → [Bj ]P (8)

Note that these axioms do not resemble successor state axioms. They nevertheless
validate the same regression principle as in Reiter’s framework, as it will be shown in
the sequel.
A point that bears noting is that our representation indeed counts as a solution to
the frame problem: the sets ; and S are both “small” (in the sense that they are much
smaller than card (PRP ) × card (ACT )), and contain no frame axioms.
Now we turn to an important result:
T HEOREM 4. — Let ; and S be obtained from given sets Poss(A), Cause + (P ),
Cause − (P ), Cond + (A, P ) and Cond − (A, P ). Then the following equivalences are
logical consequences of S in DLAP ; .
1) [A]P ↔ ¬Poss(A) ∨ P , if A6;P and A6;¬P ;
2) [A]P ↔ ¬Poss(A) ∨ (P ∧ ¬Cond − (A, P )), if A6;P and A;¬P ;
3) [A]P ↔ ¬Poss(A) ∨ Cond + (A, P ) ∨ P , if A;P and A6;¬P ;
4) [A]P ↔ ¬Poss(A) ∨ Cond + (A, P ) ∨ (P ∧ ¬Cond − (A, P )), if A;P and
A;¬P .
 Regression in Modal Logic 177

P ROOF . —
Proving (1):
(→): We are about to prove [A]P ∧ ¬P → ¬Poss(A).

1. ¬P → [A]¬P , from the hypothesis A6;P

2. [A]P ∧ ¬P → [A]P ∧ [A]¬P , from 1. by classical logic

3. [A]P ∧ [A]¬P → [A]⊥, by K and classical logic

4. [A]P ∧ ¬P → [A](P ∧ ¬P ), from 2. and 3. by syllogism

5. [A](P ∧ ¬P ) → [A]⊥, by classical logic

6. [A]P ∧ ¬P → [A]⊥, by syllogism on 4. and 5.

7. [A]⊥ → ¬Poss(A), from global axiom (4)

8. [A]P ∧ [A]¬P → ¬Poss(A), from 3. and 7. by classical logic

9. [A]P ∧ ¬P → ¬Poss(A), from 2. and 8. by classical logic

(←): We now prove ¬Poss(A) ∨ P → [A]P .

1. P → [A]P , from the hypothesis A6;¬P

2. ¬Poss(A) → [A]⊥, from global axiom (4)

3. [A]⊥ → [A]P , by K and classical logic

4. ¬Poss(A) → [A]P , from 2. and 3. by classical logic

5. ¬Poss(A) ∨ P → [A]P , from 1. and 4. by classical logic

Proving (2):
(→): Let’s show [A]P ∧ ¬P → ¬Poss(A) and [A]P ∧ Cond − (A, P ) → ¬Poss(A).

1. ¬P → [A]¬P , from the hypothesis A6;P

2. [A]P ∧ ¬P → [A]P ∧ [A]¬P , from 1. by classical logic

3. [A]P ∧ [A]¬P → [A]⊥, by K and classical logic

4. [A]⊥ → ¬Poss(A), from global axiom (4)

5. [A]P ∧ [A]¬P → ¬Poss(A), from 3. and 4. by classical logic

6. [A]P ∧ ¬P → ¬Poss(A), from 2. and 5. by classical logic

178 Journal of Applied Non-Classical Logics. Volume 13 – n◦ 2/2003

7. Cond − (A, P ) → [A]¬P , by global axiom (7)

8. [A]P ∧ Cond − (A, P ) → [A]P ∧ [A]¬P , from 7. by classical logic

9. [A]P ∧ Cond − (A, P ) → [A]⊥, from 8. and 3. by classical logic

10. [A]P ∧ Cond − (A, P ) → ¬Poss(A), from 9. and 4. by classical logic

(←): We are going to prove ¬Poss(A) ∨ (P ∧ ¬Cond − (A, P )) → [A]P .

1. ¬Poss(A) → [A]⊥, from global axiom (4)

2. [A]⊥ → [A]P , by K and classical logic

3. ¬Poss(A) → [A]P , from 1. and 2. by classical logic

4. (P ∧ ¬Cond − (A, P )) → [A]P , from global axiom (8)

5. ¬Poss(A) ∨ (P ∧ ¬Cond − (A, P )) → [A]P , from 3. and 4. by classical logic

Proving (3):
(→): We will prove [A]P ∧ ¬Cond + (A, P ) ∧ ¬P → ¬Poss(A).

1. ¬Cond + (A, P ) ∧ ¬P → [A]¬P , by global axiom (6)

2. [A]P ∧ ¬Cond + (A, P ) ∧ ¬P → [A]P ∧ [A]¬P , from 1. by classical logic

3. [A]P ∧ [A]¬P → [A]⊥, by K and classical logic

4. [A]P ∧ ¬Cond + (A, P ) ∧ ¬P → [A]⊥, from 2. and 3. by classical logic

5. [A]⊥ → ¬Poss(A), from global axiom (4)

6. [A]P ∧ ¬Cond + (A, P ) ∧ ¬P → ¬Poss(A), from 4. and 5. by classical logic

(←): We are about to prove ¬Poss(A) ∨ Cond + (A, P ) ∨ P → [A]P

1. ¬Poss(A) → [A]⊥, from global axiom (4)

2. [A]⊥ → [A]P , by K and classical logic

3. ¬Poss(A) → [A]P , from 1. and 2. by classical logic

4. P → [A]P , by hypothesis A6;¬P

5. Cond + (A, P ) → [A]P , from global axiom (5)

6. ¬Poss(A) ∨ Cond + (A, P ) ∨ P → [A]P , from 3., 4. and 5. by classical logic

 Regression in Modal Logic 179

Proving (4):
(→): We prove [A]P ∧ ¬Cond + (A, P ) ∧ ¬(P ∧ ¬Cond − (A, P )) → ¬Poss(A)

1. ¬Cond + (A, P ) ∧ ¬P → [A]¬P , from global axiom (6)

2. [A]P ∧ ¬Cond + (A, P ) ∧ ¬P → [A]P ∧ [A]¬P , from 1. by classical logic

3. Cond − (A, P ) → [A]¬P , by global axiom (7)

4. [A]P ∧ ¬Cond + (A, P ) ∧ Cond − (A, P ) → [A]P ∧ ¬Cond + (A, P ) ∧ [A]¬P ,

from 3. by classical logic

5. [A]P ∧ ¬Cond + (A, P ) ∧ [A]¬P → [A]P ∧ [A]¬P , by classical logic

6. [A]P ∧ ¬Cond + (A, P ) ∧ Cond − (A, P ) → [A]P ∧ [A]¬P , from 4. and 5. by

classical logic

7. [A]P ∧ ¬Cond + (A, P ) ∧ ¬P ∨ [A]P ∧ ¬Cond + (A, P ) ∧ Cond − (A, P ) →

[A]P ∧ [A]¬P , from 2. and 4. by classical logic

8. [A]P ∧ ¬Cond + (A, P ) ∧ ¬(P ∧ ¬Cond − (A, P )) → [A]P ∧ [A]¬P , from 7.

by classical logic

9. [A]P ∧ [A]¬P → [A]⊥, by K and classical logic

10. [A]P ∧ ¬Cond + (A, P ) ∧ ¬(P ∧ ¬Cond − (A, P )) → [A]⊥, from 8. and 9. by
classical logic

11. [A]⊥ → ¬Poss(A), from global axiom (4)

12. [A]P ∧ ¬Cond + (A, P ) ∧ ¬(P ∧ ¬Cond − (A, P )) → ¬Poss(A), from 10.
and 11. by classical logic

(←): We will prove ¬Poss(A) ∨ Cond + (A, P ) ∨ (P ∧ ¬Cond − (A, P )) → [A]P

1. ¬Poss(A) → [A]⊥, from global axiom (4)

2. [A]⊥ → [A]P , by K and classical logic

3. ¬Poss(A) → [A]P , from 1. and 2. by classical logic

4. Cond + (A, P ) → [A]P , from global axiom (5)

5. (P ∧ ¬Cond − (A, P )) → [A]P , by global axiom (8)

6. ¬Poss(A)∨Cond + (A, P ) ∨(P ∧¬Cond − (A, P )) → [A]P , from 3., 4. and 5.

by classical logic
180 Journal of Applied Non-Classical Logics. Volume 13 – n◦ 2/2003

8. Regression in DLAP ;

Here is a regression algorithm for DLAP ; . Suppose Φ is a complex formula

without quantification and equality, such as HasGun → [load][shoot]¬Alive. Let us
consider Cond (A, L) = Cond + (A, P ) if L = P , and Cond (A, L) = Cond − (A, P )
if L = ¬P .
A LGORITHM 5 (R EGRESSION WITH DEPENDENCE ). —

input:
a formula without variables Φ.
Poss(A), Cause + (P ), Cause − (P ), Cond + (A, P ) and Cond − (A, P ).
output: a classical formula REG(Φ).

begin
while Φ is not classical
put Φ in normal form
choose some subformula [A]L, where L is a literal
case A6;L and A6;¬L then replace [A]L by ¬Poss(A) ∨ L
case A6;L and A;¬L then
replace [A]L by ¬Poss(A) ∨ (L ∧ ¬Cond (A, ¬L))
case A;L and A6;¬L then replace [A]L by ¬Poss(A)∨Cond (A, L)∨L
case A;L and A;¬L then
replace [A]L by ¬Poss(A) ∨ Cond (A, L) ∨ (L ∧ ¬Cond (A, ¬L))
end

In our example, the regression of [shoot]¬Alive is ¬HasGun ∨ Loaded ∨ ¬Alive.

Hence the result of the regression step is HasGun → [load](¬HasGun ∨ Loaded ∨
¬Alive). Putting this into normal form using (3) we obtain the formula HasGun →
([load]¬HasGun∨[load]Loaded∨[load]¬Alive). The regression of [load]¬HasGun is
¬HasGun, that of [load]Loaded is >, and that of [load]¬Alive is ¬Alive. We therefore
obtain HasGun → (¬HasGun ∨ > ∨ ¬Alive), which is valid in classical propositional
logic.
T HEOREM 6 (D ECIDABILITY, SOUNDNESS AND COMPLETENESS ). — Suppose
S and ; obtained from Poss(A), Cause + (P ), Cause − (P ), Cond + (A, P ) and
Cond − (A, P ) as described in Section 7.2. Let Φ be an input formula without quanti-
fiers, action variables, and equality. Then, Algorithm 5 terminates returning a classi-
cal formula φ and S|=DLAP ;Φ ↔ φ.
P ROOF . — Let Φ be an input formula. Termination is straightforward, as each step
of the algorithm eliminates exactly one modal operator. Soundness and completeness
 Regression in Modal Logic 181

are also immediate: after putting formula Φ in normal form, it will be made of con-
junctions/disjunctions of modal subformulas. In this case, the equivalence between Φ
and φ follows from the ones given in theorem 4 together with the rule of substitution
of equivalences (which is valid in DLAP ; ).
For our example, this means that HasGun → [load][shoot]¬Alive can be deduced
with our action theory S and dependence relation ; because its regression is valid in
classical logic.
Hence, modulo the equality predicate, we obtain the same result as for Reiter’s
regression algorithm in the case of our example. This generalizes: a close look at the
two algorithms shows that if both our S and ; and Reiter’s R are obtained from the
same Poss(A), Cause + (P ), Cause − (P ), Cond + (A, P ), Cond − (A, P ), then the
results are logically equivalent.
It follows thus that whenever Poss(A), Cause + (P ), Cause − (P ), Cond + (A, P ),
Cond − (A, P ) are given, and the completeness assumptions can be made, then Re-
iter’s formulation in terms of successor state axioms and ours in terms of effect axioms
and dependence do the same job in their respective logical basis:
C OROLLARY 7. — Let the sets Poss(A), Cause + (P ), Cause − (P ), Cond + (A, P ),
Cond − (A, P ) be given. Let R be a Reiter theory obtained from them as described
in Section 4. Let ; and S be obtained from them as described in Section 7.2. Let
Φ be a complex formula without quantification and equality. Then R |= RT R Φ iff
S|=DLAP ;Φ.

9. The frame problem for knowledge

9.1. Sensing actions and knowledge

Reiter’s framework does not account for actions which have no effect on the “real”
world, but only on the agents’ knowledge. Such actions are close to test actions of
dynamic logic.
In order to express the effects of sensing actions we need a modal operator of
knowledge 2. The logic of 2 is S5. The dual of 2 is noted 3.
The extension of Reiter’s solution to knowledge and sensing actions has been stud-
ied by Scherl and Levesque [SCH 93]. They make some hypotheses about actions and
their perception by the agent that permit to simplify the theory.

9.1.1. Public action

First, they suppose that the agent perceives action occurrences completely and
correctly. For example whenever shooting takes place the agent is aware of that, and
whenever the agent believes shooting has taken place then indeed such an action has
occurred. (One might imagine that action occurrences are publicly announced to all
agents.)
182 Journal of Applied Non-Classical Logics. Volume 13 – n◦ 2/2003

9.1.2. Action laws known

Second, they suppose that the agent knows the laws governing the actions. Hence
the agent knows that after strangling the effect always is ¬Alive, etc.

9.1.3. Non-informative actions

We finally make a third hypothesis that is not made by Scherl and Levesque, but
which simplifies exposition without too much loss of generality [HER 00]. We shall
suppose henceforth that all actions are non-informative. Non-informative actions are
actions which are not observed by the agent beyond their mere occurrence. Upon
learning that such an action has occurred the agent updates his belief state: he com-
putes the new belief state from the previous one and his knowledge about the action
laws. Hence the new belief state neither depends on the state of the world before the
action occurrence, nor on the state of the world after the action occurrence.
In our example the shoot action is non informative. If the agent learns that the
shooting action has been executed then he does not learn whether the victim died or
not: if both Loaded and ¬Loaded were possible for the agent before, then afterwards
he envisages both possible outcomes. Nevertheless, by learning that shoot occurred
the agent learns that HasGun was true before the action.
Clearly, the action of observing the outcome of the shoot action is informative: the
new belief state depends on the truth value of ¬Alive in the real world. Other examples
of informative actions are that of looking up a phone number, testing if a proposition
is true, telling whether a proposition is true, etc.
Nevertheless, the agent is not disconnected from the world: he may learn that some
proposition is true (i.e. that some action of observing that some proposition has some
value has occurred). For example, when he learns that it has been observed that the
victim is dead (i.e. he learns that the action of observing ¬Alive has been executed)
then he is able to update his belief state accordingly. Indeed, the observe actions are
non-informative according to our definition: when the agent learns that φ has been
observed then he is able to update his belief state accordingly, and there is no need
to further observe the world. Other examples of non-informative actions are that of
learning that the phone number of another agent is N , testing that a proposition is
true (in the sense of dynamic logic tests), telling that a proposition is true, etc.

9.1.4. A successor state axiom

Under the hypotheses we have made, the following logical axiom is reasonable:

[A]2Φ ↔ ([A]⊥ ∨ 2[A]Φ)

From the left to the right, this corresponds to a “no forgetting” principle, while the
right-to-left direction expresses a “no learning” principle.
 Regression in Modal Logic 183

Such an axiom has been called a successor state axiom by Scherl and Levesque. 4
It permits to solve what they have called the frame problem for knowledge.
First, note that because actions are supposed to be deterministic, this axiom allows
to deduce
[A]3Φ ↔ ([A]⊥ ∨ 3¬[A]¬Φ).

Now these two principles enable regression by allowing for the elimination of 2
and 3 operators from the scope of action operators. When all such epistemic oper-
ators have been moved outward, Reiter’s regression can be applied to the remaining
non-epistemic formula, resulting in a modality-free formula of classical propositional
logic. As a whole, the resulting formula only contains epistemic operators, but no
action operators.
To sum it up, to establish whether a complex formula Φ follows from a domain
description amounts to
– move all [A] operators inwards, then
– eliminate all [A] operators by regression, and finally
– check whether the resulting formula Ψ is a theorem of S5.

10. Concluding remarks

In this paper we have presented a purely propositional framework for reasoning

about actions in modal logic within which Reiter’s regression technique can be ap-
plied. We have thus shown that regression does not necessarily build on successor
state axioms as in Reiter’s original theory, which involves quantification.
We have also seen how the ideas here developed could be extended and applied in
reasoning about knowledge.
As we have presented it here, Reiter’s solution is very constrained. In particular
actions must be deterministic and without indirect effects.
Reiter has proposed [REI 01] to implement nondeterministic actions by means of
an operator of nondeterministic composition of deterministic atomic actions similar to
that of dynamic logic. For example, the action toss of tossing a coin can be thought of
as the nondeterministic choice tossHeads∪tossTails between tossHeads and tossTails,
whose respective effects are Heads and Tails. [tossHeads ∪ tossTails]Φ is defined to
be an abbreviation of [tossHeads]Φ ∧ [tossTails]Φ. Such a solution transfers straight-
forwardly to our modal logic. But our framework also offers a more straightforward
way of dealing with actions with indeterminate effects: we can drop the hypothesis
that for every A ∈ Cause + (P ) the condition Cond + (A, P ) is defined. For example,

4. Their successor state axiom contains supplementary conditions in order to account for the
informative part of actions.
184 Journal of Applied Non-Classical Logics. Volume 13 – n◦ 2/2003

although Cause + (Heads) = Cause + (Tails) = {toss}, there is no way of stating the
exact conditions when heads or tails results from tossing.
Reiter’s solution supposes that domain descriptions only contain executability and
effect laws. Thus it does not allow for static laws such as Walking → Alive. Such
laws augment the effects of the shoot action: shooting not only has the (direct) ef-
fect ¬Alive, but also the (indirect) effect ¬Walking. Reiter and Lin [LIN 94] have
proposed to “compile away” static laws in a mechanical way into effect laws (see
also [MCI 98]). Again, this transfers straightforwardly to our modal logic. Never-
theless, the most challenging continuation of our work is the direct integration of so-
called state constraints into the framework (instead of compiling them away as done
by Lin and Reiter). But things get much harder in this case, all the more in [CAS 02]
we have claimed that up to now there is no satisfactory framework allowing for actions
with both indirect and indeterminate effects.
We plan to pursue future works analyzing to at what extent the results here pre-
sented could be generalized to Lin’s [LIN 95, LIN 96] approach in the case of stratified
action theories.

Acknowledgements

Ivan Varzinczak has been supported by a fellowship from the Federative Republic
of Brazil. Grant: BEX 1389/01-7 (CAPES).

11. References

[CAS 99] C ASTILHO M. A., G ASQUET O., H ERZIG A., “Formalizing Action and Change in
Modal Logic I: the frame problem”, J. of Logic and Computation, vol. 9, num. 5, 1999,
p. 701–735.
[CAS 02] C ASTILHO M. A., H ERZIG A., VARZINCZAK I., “It depends on the context! A
decidable logic of actions and plans based on a ternary dependence relation”, Proc. of
NMR’2002, 2002.
[DEM 03] D EMOLOMBE R., “Belief change: from Situation Calculus to Modal Logic”,
B REWKA G., P EPPAS P., Eds., Proc. of the Workshop on Nonmonotonic Reasoning, Ac-
tion and Change, 2003.
[FIT 83] F ITTING M., Proof Methods for Modal and Intuitionistic Logics, D. Reidel, Dor-
drecht, 1983.
[GIA 95] G IACOMO G. D., L ENZERINI M., “PDL-based framework for reasoning about ac-
tions”, G ORI M., S ODA G., Eds., Proc. of AI*IA’95, vol. 992 of LNAI, 1995, p. 103–114.
[HAN 86] H ANKS S., M C D ERMOTT D., “Default reasoning, nonmonotonic logics and the
frame problem”, Proc. of AAAI’86, 1986, p. 328–333.
[HAR 84] H AREL D., “Dynamic Logic”, G ABBAY D. M., G UENTHER F., Eds., Handbook
of Philosophical Logic, vol. 2: Extensions of Classical Logic, p. 497–604, D. Reidel
Publishing Co., Dordrecht, 1984.
 Regression in Modal Logic 185

[HER 00] H ERZIG A., L ANG J., P OLACSEK T., “A modal logic for epistemic tests”, Proc. of
ECAI’2000, Aug. 2000.
[LIN 94] L IN F., R EITER R., “State Constraints Revisited”, J. of Logic and Computation,
vol. 4, num. 5, 1994, p. 655–677.
[LIN 95] L IN F., “Embracing Causality in Specifying the Indirect Effects of Actions”, M EL -
LISH C., Ed., Proc. of IJCAI’95, Montreal, 1995, p. 1985–1991.
[LIN 96] L IN F., “Embracing Causality in Specifying the Indeterminate Effects of Actions”,
Proc. of AAAI’96, vol. 1, 1996, p. 670–676.
[MCC 69] M C C ARTHY J., H AYES P., “Some philosophical problems from the standpoint of
artificial intelligence”, Machine Intelligence, vol. 4, 1969, p. 463–502.
[MCI 98] M C I LRAITH S., “Representing Action and State Constraints in Model-Based Diag-
nosis”, Proc. of AAAI’98, Menlo Park, California, 1998, p. 43–49.
[OHL 91] O HLBACH H. J., “Semantics Based Translation Methods for Modal Logics”, J. of
Logic and Computation, vol. 1, num. 5, 1991, p. 691–746.
[OHL 93] O HLBACH H. J., “Translation Methods for Non-Classical Logics – An Overview”,
J. of the Interest Group in Pure and Applied Logics, vol. 1, num. 1, 1993, p. 69–90.
[REI 91] R EITER R., “The frame problem in the Situation Calculus: a Simple Solution (some-
times) and a Completeness Result for Goal Regression”, Artificial Intelligence and Mathe-
matical Theory of Computation, Papers in Honor of John McCarthy, p. 359–380, Academic
Press, 1991.
[REI 01] R EITER R., Knowledge in Action: Logical Foundations for Specifying and Imple-
menting Dynamical Systems, The MIT Press, Cambridge, MA, 2001.
[SCH 90] S CHUBERT L. K., “Monotonic solution of the frame problem in the situation calcu-
lus: an efficient method for worlds with fully specified actions”, Knowledge Representation
and Defeasible Reasoning, 1990, p. 23–67.
[SCH 93] S CHERL R. B., L EVESQUE H. J., “The frame problem and knowledge producing
actions”, Proc. of AAAI’93, Washington, DC, 1993, p. 689–697.
[ZHA 01] Z HANG D., F OO N. Y., “EPDL: A Logic for Causal Reasoning”, Proc. of IJ-
CAI’2001, 2001, p. 131–138.