Scaling up Probabilistic Inference in Linear and Non-linear Hybrid

Domains by Leveraging Knowledge Compilation

Anton R. Fuxjaeger

1

and Vaishak Belle

1,2

1

University of Edinburgh, U.K.

2

Alan Turing Institute, U.K.

Keywords:

Weighted Model Integration, Probabilistic Inference, Knowledge Compilation, Sentential Decision Diagrams,

Satisﬁability Modulo Theories.

Abstract:

Weighted model integration (WMI) extends weighted model counting (WMC) in providing a computational

abstraction for probabilistic inference in mixed discrete-continuous domains. WMC has emerged as an as-

sembly language for state-of-the-art reasoning in Bayesian networks, factor graphs, probabilistic programs

and probabilistic databases. In this regard, WMI shows immense promise to be much more widely applicable,

especially as many real-world applications involve attribute and feature spaces that are continuous and mixed.

Nonetheless, state-of-the-art tools for WMI are limited and less mature than their propositional counterparts.

In this work, we propose a new implementation regime that leverages propositional knowledge compilation for

scaling up inference. In particular, we use sentential decision diagrams, a tractable representation of Boolean

functions, as the underlying model counting and model enumeration scheme. Our regime performs compet-

itively to state-of-the-art WMI systems but is also shown to handle a speciﬁc class of non-linear constraints

over non-linear potentials.

1 INTRODUCTION

Weighted model counting (WMC) is a basic reason-

ing task on propositional knowledge bases. It ex-

tends the model counting task, or #SAT, which is

to count the number of satisfying assignments to a

given propositional formula (Biere et al., 2009). In

WMC, one accords a weight to every model and com-

putes the sum of the weights of all models. The

weight of a model is often factorized into weights

of assignments to individual variables. WMC has

emerged as an assembly language for numerous for-

malisms, providing state-of-the-art probabilistic rea-

soning for Bayesian networks (Chavira and Dar-

wiche, 2008), factor graphs (Choi et al., 2013),

probabilistic programs (Fierens et al., 2015), and

probabilistic databases (Suciu et al., 2011). Ex-

act WMC solvers are based on knowledge compi-

lation (Darwiche, 2004; Muise et al., 2012) or ex-

haustive DPLL search (Sang et al., 2005). These

successes have been primarily enabled by the devel-

opment of efﬁcient data structures, e.g., arithmetic

circuits (ACs), for representing Boolean theories, to-

gether with fast model enumeration strategies. In par-

ticular, the development of ACs has enabled a number

of developments beyond inference, such as parame-

ter and structure learning (Bekker et al., 2015; Liang

et al., 2017; Poon and Domingos, 2011; Kisa et al.,

2014; Poon and Domingos, 2011). Finally, having

a data structure in hand means that multiple queries

can be evaluated efﬁciently: that is, exhaustive search

need not be re-run for each query.

However, WMC is limited to discrete ﬁnite-

outcome distributions only, and little was understood

about whether a suitable extension exists for contin-

uous and discrete-continuous random variables until

recently. The framework of weighted model integra-

tion (WMI) (Belle et al., 2015) extended the usual

WMC setting by allowing real-valued variables over

symbolic weight functions, as opposed to purely nu-

meric weights in WMC. The key idea is to use formu-

las involving real-valued variables to deﬁne a hyper-

rectangle or a hyper-rhombus, or in general, any arbi-

trary region of the event space of a continuous ran-

dom variable, and use the symbolic weights to de-

ﬁne the corresponding density function for that re-

gion.WMC is based on propositional SAT technol-

ogy and, by extension, WMI is based on satisﬁability

modulo theories (SMT), which enable us to, for ex-

ample, reason about the satisﬁability of linear con-

Fuxjaeger, A. and Belle, V.

Scaling up Probabilistic Inference in Linear and Non-linear Hybrid Domains by Leveraging Knowledge Compilation.

DOI: 10.5220/0008896003470355

In Proceedings of the 12th International Conference on Agents and Artiﬁcial Intelligence (ICAART 2020) - Volume 2, pages 347-355

ISBN: 978-989-758-395-7; ISSN: 2184-433X

Copyright

c

2022 by SCITEPRESS – Science and Technology Publications, Lda. All rights reserved

347

straints over the reals (Barrett et al., 2009). Thus,

for every assignment to the Boolean and continuous

variables, the WMI problem deﬁnes a density. The

WMI for a knowledge base (KB) ∆ is computed by

integrating these densities over the domain of so-

lutions to ∆, which is a mixed discrete-continuous

space, yielding the value for a probabilistic query.

The approach is closely related to the mixture-of-

polynomials density estimation for hybrid Bayesian

networks (Shenoy and West, 2011). Applications

of WMI (and closely related formulations) for prob-

abilistic graphical modelling and probabilistic pro-

gramming tasks have also been emerging (Chistikov

et al., 2017; Albarghouthi et al., 2017; Morettin et al.,

2017).

Given the popularity of WMC, WMI shows im-

mense promise to be much more widely applicable,

especially as many real-world applications, includ-

ing time-series models, involve attribute and feature

spaces that are continuous and mixed. However, state-

of-the-art tools for WMI are limited and signiﬁcantly

less mature than their propositional counterparts. Ini-

tial developments on WMI (Belle et al., 2015) were

based on the so-called block-clause strategy, which

naively enumerates the models of a LR A (linear real

arithmetic) theory and is impractical on all but small

problems. Recently, a solver based on predicate ab-

straction was introduced by (Morettin et al., 2017)

with strong performance, but since no explicit circuit

is constructed, it is not clear how tasks like parameter

learning can be realized. Following that development,

(Kolb et al., 2018) proposed the use of extended alge-

braic decision diagrams (Sanner et al., 2012), an ex-

tension of algebraic decision diagrams (Bahar et al.,

1997), as a compilation language for WMI. They also

perform comparably to (Morettin et al., 2017).

However, while this progress is noteworthy, there

are still many signiﬁcant differences to the body of

work on propositional circuit languages. For exam-

ple, properties such as canonicity have received con-

siderable attention for these latter languages (Van den

Broeck and Darwiche, 2015). Many of these lan-

guages allow (weighted) model counting to be com-

puted in time linear in the size of the obtained circuit.

To take advantage of these results, in this work we re-

visit the problem of how to leverage propositional cir-

cuit languages for WMI more carefully and develop a

generic implementation regime to that end. In partic-

ular, we leverage sentential decision diagrams (SDDs)

(Darwiche, 2011) via abstraction. SDDs are tractable

circuit representations that are at least as succinct

as ordered binary decision diagrams (OBDDs) (Dar-

wiche, 2011). Both of these support querying such as

model counting (MC) and model enumeration (ME)

in time linear in the size of the obtained circuit. (We

use the term querying to mean both probabilistic con-

ditional queries as well as weighted model count-

ing because the latter simply corresponds to the case

where the query is true.) Because of SDDs having

such desirable properties, several papers have dealt

with more involved issues, such as learning the struc-

ture from data directly (Bekker et al., 2015; Liang

et al., 2017) and thus learning the structure of the un-

derlying graphical model.

In essence, our implementation regime uses SDDs

as the underlying querying language for WMI in or-

der to perform tractable and scalable probabilistic in-

ference in hybrid domains. The regime neatly sep-

arates the model enumeration from the integration,

which is demonstrated by allowing a choice of two

integration schemes. The ﬁrst is a provably efﬁcient

and exact integration approach for polynomial den-

sities (De Loera et al., 2004; Baldoni et al., 2011;

De Loera et al., 2011) and the second is an unmodi-

ﬁed integration library available in the programming

language platform (Python in our case). The results

obtained are very promising with regards to the em-

pirical behaviour: we perform competitively to the

existing state-of-the-art WMI solver (Morettin et al.,

2017). But perhaps most signiﬁcantly, owing to the

generic nature of our regime, we can scale the same

approach to non-linear constraints, with possibly non-

linear potentials.

2 BACKGROUND

Probabilistic Graphical Models. Throughout this

paper we will refer to Boolean and continuous ran-

dom variables as B

j

and X

i

respectively for some ﬁ-

nite j > 0,i > 0. Lower case letters, b

j

∈ {0,1} and

x

i

∈ R, will represent the instantiations of these vari-

ables. Bold upper case letters will denote sets of vari-

ables and bold lower case letters will denote their

instantiations. We are broadly interested in proba-

bilistic models, deﬁned on B and X. That is, let

(b,x) = (b

1

,b

2

,. .. ,b

m

,x

1

,x

2

,. .. ,x

n

) be one element

in the probability space {0,1}

m

∗ R

n

, denoting a par-

ticular assignment to the values in the respective do-

mains. A graphical model can then be used to de-

scribe dependencies between the variables and deﬁne

a joint density function of those variables compactly.

The graphical model we will consider in this paper

are Markov networks, which are undirected models.

(Directed models can be considered too (Chavira and

Darwiche, 2008), but are ignored for the sake of sim-

plicity.)

ICAART 2020 - 12th International Conference on Agents and Artiﬁcial Intelligence

348

Logical Background. Propositional satisﬁability

(SAT) is the of determining if a given formula in

propositional logic can be satisﬁed by an assignment

(, where a satisfying assignment has to be provided

as proof for a formula being satisﬁable). An instance

of satisﬁability modulo theory (SMT) (Biere et al.,

2009) is a generalization of classical SAT in allow-

ing ﬁrst-order formulas with respect to some decid-

able background theory. For example, LR A is under-

stood here as quantiﬁer-free linear arithmetic formu-

las over the reals and the corresponding background

theory is the fragment of ﬁrst-order logic over the

signature (0,1,+,≤), restricting the interpretation of

these symbols to standard arithmetic.

In this work, we will consider two different back-

ground theories: quantiﬁer-free linear (LR A) and

non-linear (N R A) arithmetic over the reals. A prob-

lem instance (input) to our WMI solver is then a for-

mula with respect to one of those background theories

in combination with propositional logic for which sat-

isfaction is deﬁned in an obvious way (Barrett et al.,

2009). Such an instance is referred to as a hybrid

knowledge base (HKB).

Weighted Model Counting. Weighed model

counting (WMC) (Chavira and Darwiche, 2008) is a

strict generalization of model counting (Biere et al.,

2009). In WMC, each model of a given propositional

knowledge base (PKB) Γ has an associated weight

and we are interested in computing the sum of

the weights that correspond to models that satisfy

Γ. (As is convention, the underlying propositional

language and propositional letters are left implicit.

We often refer to the set of literals L to mean the set

of all propositional atoms as well as their negations

constructed from the propositions mentioned in Γ.)

In order to create an instance of the WMC prob-

lem given a PKB Γ and literals L, we deﬁne a weight

function wf : L → R

≥0

mapping the literals to non-

negative, numeric weights. We can then use the lit-

erals of a given model m to deﬁne the weight of that

model as well as the weighted model count as follows:

Deﬁnition 1. Given a PKB Γ over literals L (con-

structed from Boolean variables B) and weight func-

tion wf : L → R

≥0

, we deﬁne the weight of a model

as:

WEIGHT(m,wf ) =

∏

l∈m

wf (l) (1)

Further we deﬁne the weighted model count (WMC)

as:

WMC(Γ,wf ) =

∑

m|=Γ

WEIGHT(m,wf ) (2)

It can be shown that WMC can be used to cal-

culate probabilities of a given graphical model N

by means of a suitable encoding (Chavira and Dar-

wiche, 2008). In particular, conditional probabilities

can be calculated using: Pr

N

(q|e) =

WMC(Γ∧q∧e,wf )

WMC(Γ∧e,wf )

for some evidence e and query q, where e,q are PKBs

as well, deﬁned from B.

Weighted Model Integration. While WMC is very

powerful as an inference tool, it suffers from the in-

herent limitation of only admitting inference in dis-

crete probability distributions. This is due to its un-

derlying theory in enumerating all models (or expand-

ing the complete network polynomial), which is ex-

ponential in the number of variables, but still ﬁnite

and countable in the discrete case. For the continuous

case, we need to ﬁnd a language to reason about the

uncountable event spaces, as well as represent density

functions. WMI (Belle et al., 2015) was proposed as

a strict generalization of WMC for hybrid domains,

with the idea of annotating a SMT theory with poly-

nomial weights.

Deﬁnition 2. (Belle et al., 2015) Suppose ∆ is a HKB

over Boolean and real variables B and X, and literals

L. Suppose wf : L → EXPR(X), where EXPR(X) are

expressions over X. Then we deﬁne WMI as:

WMI(∆,wf ) =

∑

m|=∆

−

VOL(m,wf ) (3)

where:

VOL(m,wf ) =

Z

{l

+

:l∈m}

WEIGHT(m,wf )dX (4)

and WEIGHT is deﬁned as described in Def 1.

Intuitively the WMI of an SMT theory ∆ is de-

ﬁned in terms of the models of its propositional ab-

straction ∆

−

. For each such model we compute its

volume, that is, we integrate the WEIGHT-values of

the literals that are true in the model. The interval of

the integral is deﬁned in terms of the reﬁnement of the

literals. The weight function wf is to be seen as map-

ping an expression e to its density function, which is

usually another expression mentioning the variables

appearing in e. Conditional probabilities can be cal-

culated as before.

Sentential Decision Diagram. Sentential decision

diagrams (SDDs) were ﬁrst introduced in (Darwiche,

2011) and are graphical representations of propo-

sitional knowledge bases. SDDs are shown to be

a strict subset of deterministic decomposable nega-

tion normal form (d-DNNF), a popular representation

for probabilistic reasoning applications (Chavira and

Scaling up Probabilistic Inference in Linear and Non-linear Hybrid Domains by Leveraging Knowledge Compilation

349

Darwiche, 2008) due to their desirable properties. De-

composability and determinism ensure tractable prob-

abilistic (and logical) inference, as they enable MAP

queries in Markov networks. SDDs however satisfy

two even stronger properties found in ordered binary

decision diagrams (OBDD), namely structured de-

composability and strong determinism. Indeed, (Dar-

wiche, 2011) showed that they are strict supersets

of OBDDs as well, inheriting their key properties:

canonicity and a polynomial time support for Boolean

combination. Finally SDD’s also come with an upper

bound on their size in terms of tree-width. In the in-

terest of space, we will not be able to discuss SDD

properties in detail. However, we refer the reader to

the original paper (Darwiche, 2011) for an in-depth

study of SDDs and the central results of SDDs that

we appeal to.

3 METHOD

Over the past few years there have been several papers

on exact probabilistic inference (Morettin et al., 2017;

Sanner et al., 2012; Kolb et al., 2018) using the for-

mulation of WMI. What we propose in this section is

a novel formulation of doing weighted model integra-

tion by using SDDs as the underlying model counting,

enumeration and querying language. Here predicate

abstraction and knowledge compilation enable us to

compile the abstracted PKB into an SDD, which has

the desirable property of a fully parallelisable poly-

time model-enumeration algorithm. Recall that poly-

time here refers to the complexity of the algorithm

with respect to the size of the tree (SDD) (Darwiche

and Marquis, 2002).

In practice, computing the probability of a given

query for some evidence consists of calculating the

WMI of two separate but related HKBs. That is, we

have to compute the WMI of a given HKB ∆ con-

joined with some evidence e and the query q, divid-

ing it by the WMI of ∆ conjoined with the evidence

e. This formulation introduced by (Belle et al., 2015)

and explained in more detail in Section 2, can be writ-

ten as:

Pr

∆

(q|e) =

WMI(∆ ∧ e ∧ q)

WMI(∆ ∧ e)

(5)

We will give a quick overview of the whole

pipeline for computing the WMI value of a given

KB, before discussing in detail the individual com-

putational steps.

Step 1: Predicate Abstraction

Step 2: Knowledge Compilation

Step 3: Model Enumeration

Step 4: Integration

'Separable' HKB

Propositional KB

SDD

Satisfying Models

Weight Function

Prop. Refinements

Problem Instance

Prop. Logic

WMI Value

WMI-SDD

Figure 1: Pictorial depiction of the proposed pipeline for

WMI.

3.1 WMI-SDD: The Pipeline

As a basis for performing probabilistic inference, we

ﬁrst have to be able to calculate the WMI of a given

HKB ∆ with corresponding weight function wf . As

we are interested in doing so by using SDDs as a

query language, the WMI breaks down into a se-

quence of sub-computations depicted as the WMI-

SDD pipeline in Figure 1.

Input/Outputs of the Pipeline. The input of the

pipeline is composed of two things: the HKB with re-

spect to some background theory (eg. LR A,N R A)

on the one hand and the weight function on the other.

Here, atoms are deﬁned as usual for the respective

language (Barrett et al., 2009) and can be under-

stood as functions that cannot be broken down fur-

ther into a conjunction, disjunction or negation of

smaller expressions. This means that a HKB of the

form ((X

1

< 3) ∧ (X

1

> 1)) should be abstracted as

(B

1

∧ B

2

) with B

+

1

= (X

1

< 3) and B

+

2

= (X

1

> 1),

rather than B

0

with B

+

0

= (X

1

< 3) ∧ (X

1

> 1).

The ﬁrst step is to arrange atoms in a form that

we call ‘separable’. The corresponding background

theory determines whether a correct rewriting of for-

mulas is possible to satisfy this condition:

Deﬁnition 3. A given HKB ∆ satisﬁes the condi-

tion separable if every atom within the formula can

be rewritten in one of the following forms: X

1

<

d(A), d(A) < X

1

, X

1

≤ d(A), d(A) ≤ X

1

or d(A) ≤

ICAART 2020 - 12th International Conference on Agents and Artiﬁcial Intelligence

350

X

1

∧ X

1

≤ d(A) where d(A) is any term over A ⊆

VARS − {X

1

}, with VARS being the set of all vari-

ables (Boolean and continuous) that appear in the

atom. That is, by construction, X

1

/∈ A for any given

variable X

1

∈ VARS. Such a variable X

1

is then called

the leading variable (leadVar).

For some background theories, this conversion is

immediate. In a LR A formula ∆

LR A

, any given atom

can be rewritten as an inequality or equality where we

have a single variable on one side and a linear function

on the other side, such as (X

1

< 3 + X

2

). But this is

not a given for HKBs with background theory N R A.

For example, (3 < 2 ∗ X

1

+ X

2

2

) can be rewritten as

(X

1

< 3/2 − 1/2 ∗ X

2

2

) for X

1

and therefore satisﬁes

the condition. However, the atom (3 < X

4

1

− 3 ∗ X

2

1

)

cannot be rewritten in a similar manner and thus does

not satisfy the condition.

Considering the motivation of performing proba-

bilistic inference, where we deal with evidence and

queries in addition to an HKBs, as discussed in Sec-

tion 2, we note that all elements of {∆,q,e} have to

fulﬁl the separability condition. As queries and ev-

idence are applied by means of a logical connective

with the HKB, they should generally be thought of as

HKBs themselves.

The weight function wf , on the other hand,

is only restricted by the condition that the term

WEIGHT(m,wf ) must be integratable for any given

model m. As long as this condition is met, we

can accept any arbitrary function over the variables

(Boolean and continuous) of the KB.

3.2 Step 1: Predicate Abstraction

The aim of this step in the WMI framework is

twofold. On the one hand, it is given an HKB (∆) and

is tasked to produce a PKB (∆

−

) and the correspond-

ing mapping from propositional variables to contin-

uous reﬁnements, utilizing propositional abstraction.

On the other hand, this part of the framework also re-

arranges the continuous reﬁnements such that a single

variable is separated from the rest of the equation to

one side of the inequality/equality.

On a conceptual level, the predicate abstraction

closely follows the theoretical formulation introduced

in (Belle et al., 2015). The HKB is recursively tra-

versed and every encountered atom is replaced with

a propositional variable, while the logical structure

(connectives and parentheses) of the KB is preserved.

We make use of the imposed separable property

to rewrite the individual reﬁnements into bounds for

a given variable. These bounds can easily be negated

and will be used at a later stage to construct the inter-

vals of integration for a given model. Now the pro-

cess of rewriting a single atom corresponds to sym-

bolically solving an equation for one variable and it is

implemented as an arithmetic solver. The variable we

choose to isolate from the rest of the equation (that is,

the leading variable), is determined by a variable or-

der, that in turn enforces the order of integration in a

later stage of the pipeline. For example, assume that

the chosen variable order is the usual alphabetical one

over the variable names. Then predicates are rewrit-

ten such that from all variables referenced in the atom,

the one highest up in the variable order is chosen as

the leading variable and separated from the rest of the

equation, resulting in a bound for the given variable.

This ensures that for any predicate the bound for the

leading variable does not reference any variable that

precedes it alphabetically, which in turn ensures that

the integral to be computed is deﬁned and will result

in natural number representing the volume.

Example 1. To illustrate this with an example, con-

sider the HKB ∆: ∆ = (B

0

∧ (X

1

< 3) ∧ (0 < X

1

+

X

2

)) ∨ (X

2

< 3 ∧ X

2

> 0). After abstraction we are

given the PKB ∆

−

= (B

0

∧B

1

∧B

2

)∨ (B

3

∧B

4

) where

the abstracted variables correspond to the follow-

ing atoms: B

+

1

= (X

1

< 3), B

+

2

= (0 < X

1

+ X

2

),

B

+

3

= (X

2

< 3) and B

+

4

= (X

2

> 0). As mentioned

above, we construct the order of the continuous vari-

able alphabetically, resulting in {1 : X

1

,2 : X

2

} for

the proposed example. Once the order has been con-

structed we can rewrite each predicate as a bound

for the variable appearing ﬁrst in the order: B

1

=

(X

1

< 3), B

2

= (−1 ∗ X

2

< X

1

), B

3

= (X

2

< 3)

and B

4

= (0 < X

2

). This ensures that the integral

R R

wf (X

1

,X

2

)dX

1

dX

2

computes a number for every

possible model of the KB. Considering for example

the model [B

0

,B

1

,B

2

,B

3

,B

4

], the bounds of the inte-

gral would be as follows:

R

3

0

R

3

−X

2

wf (X

1

,X

2

)dX

1

dX

2

and yields a number.

In the case of non-linear reﬁnements, the step of

rearranging the variable could give rise to new propo-

sitions, that in turn have to be added to the PKB.

Consider, for example that the predicate B with the

reﬁnement: B

+

= (4 < X

1

∗ X

2

) should be rewrit-

ten for the variable X

1

as the leading one. Now as

the variable X

2

might be negative or zero, we are

unable to simply divide both sides by X

2

but rather

have to split up the equation in the following way:

B

+

new

= (((X

2

> 0) → (4/X

2

< X

1

)) ∧ ((X

2

< 0) →

(4/X

2

> X

1

)) ∧ ((X

2

= 0) → False)) which can be

further abstracted as: B

+

new

= ((B

1

→ B

2

) ∧ (B

3

→

B

4

)∧((¬B

1

∧¬B

3

) → False)). Once created, we can

replace B with its Boolean function reﬁnement in the

PKB and add all the new predicates (B

1

,B

2

,B

3

,B

4

) to

our list of propositions.

Scaling up Probabilistic Inference in Linear and Non-linear Hybrid Domains by Leveraging Knowledge Compilation

351

3.3 Step 2: Knowledge Compilation

In this step of our pipeline, the PKB constructed in

the previous step is compiled into a canonical SDD. In

practice, we ﬁrst convert the PKB to CNF before pass-

ing it to the SDD library.

1

The library has a number of

optimizations in place, including dynamic minimiza-

tion (Choi and Darwiche, 2013). However, the algo-

rithm is still constrained by the asymptotically expo-

nential nature of the problem. In addition, it requires

the given PKB to be in CNF or DNF format. Once

the SDD is created, it is imported back into our inter-

nal data structure, which is designed for retrieving all

satisfying models of a given SDD.

3.4 Step 3: Model Enumeration

Retrieving all satisfying models of a given PKB is

a crucial part of the WMI formulation and we now

focus on this step in our pipeline. In essence, we

make use of knowledge compilation to compile the

given PKB into a data structure, which allows us to

enumerate all satisfying models in polynomial time

with respect to the size of the tree. As discussed in

the background section, SDDs are our data structures

of choice and their properties, including canonicity,

make them an appealing choice for our pipeline.

The algorithm we developed for retrieving the sat-

isfying models makes full use of the structural prop-

erties of SDDs. By recursively traversing the tree

bottom-up, models are created for each node in the

SDD with respect to the vtree node it represents.

Those models are then passed upwards in the tree

where they are combined with other branches. This is

possible due to the structured decomposability prop-

erty of the SDD data structure. It should also be noted

at this point that parallelisation of the algorithm is

possible as well due to SDDs decomposability proper-

ties. This is a highly desirable attribute when it comes

to scaling to very large theories.

3.5 Step 4: Integration

The workload of this part of the framework is to com-

pute the volume (VOL) (as introduced in Def 2) for

every satisfying model that was found in the previ-

ous step. That volume for a given model of the PKB

is computed by integrating the weight function (wf )

over the literals true at the model, where the bound of

the integral corresponds to the reﬁnement and truth

value of a given propositional variable within the

model. All such volumes are then summed together

and give the WMI value of the given HKB.

1

http://reasoning.cs.ucla.edu/sdd/

Computing a volume for a given model consists of

two parts: ﬁrstly we have to combine the reﬁnements

of predicates appropriately, creating the bounds of in-

tegration before actually integrating over the wf with

respect to the variables and bounds. As discussed in

the predicate abstraction and rewriting step, a given

predicate (that has a reﬁnement) consists of a leading

variable and a bound for the variable. Combining the

bounds into an interval is explained in Algorithm 1.

Algorithm 1 : Combining the intervals for a leadVar and

model.

1: procedure COMBINE(leadVar, predicates, model)

2: interval ← (-inf,inf)

3: for pred in predicates do

4: if pred.leadVar 6= leadVar then

5: continue

6: if model[pred.idx] == false then

7: newBound = negate(pred.bound)

8: else

9: newBound = pred.bound

10: interval = combine(interval,newBound)

11: return interval

Here the function combine combines

intervals via intersections. For exam-

ple, combine((-inf,inf), (-inf,X

1

< 3)) =

(-inf,min(inf, 3)) = (-inf,3) and combine((X

2

+

X

3

,inf), (X

2

/3∗ X

2

< X

1

,inf)) = (max(X

2

+X

3

,X

3

/3∗

X

2

),inf). This procedure is done for every variable

referenced in wf , ensuring that we have a bound of

integration for every such variable.

Naturally, not all abstracted models have to be

models of the original SMT theory. For example,

suppose that a model makes both X

0

< 5 and X

0

>

10 true, abstracted as B

1

and B

2

, then the proposi-

tional abstraction erroneously retrieves a model where

[B

1

,B

2

,. ..], and so the interval bounds would be

(10 < X

0

< 5). Clearly, then, the model should not be

considered as a model for the SMT theory and is sim-

ply disregarded. Once all the real bounds of integra-

tion are deﬁned for the given model, the next step be-

fore integrating is to enumerate all possible instantia-

tions of Boolean variables referenced in the wf . The

different integration problems are then hashed such

that the system only has to compute the integration

once, even if they appear multiple times.

When it comes to the implementation of this part

of the framework, we used two different integration

methods. We support the integration module of the

scipy python package

2

to compute the deﬁned inte-

gral for a given wf , a set of intervals and the instan-

tiations of Boolean variables. Using this package al-

lowed us to formalize the method as described above

and perform inference in non-linear domains. How-

2

https://scipy.org/

ICAART 2020 - 12th International Conference on Agents and Artiﬁcial Intelligence

352

ever, this formulation is not exact and suffers from a

slow runtime. For this reason, we also implemented

the pipeline using latte,

3

an exact integration software

that is particularly well-suited for piecewise polyno-

mial density approximations.

4 EMPIRICAL EVALUATION

Here, we evaluate the proposed framework on the

time it needs to compute the WMI of a given HKB

and wf . It is a proof-of-concept system for WMI via

SDDs. To evaluate the framework, we randomly gen-

erate problems, as described below and compare the

time to the WMI-PA framework developed in (Moret-

tin et al., 2017).

4

4.1 Problem Set Generator

A problem is generated based on 3 factors: the num-

ber of variables, the number of clauses and the per-

centage of real variables.

When generating a new Boolean atom, we simply

return a Boolean variable with the given ID, whereas

generating a real-valued atom is more intricate and

depends on the kind of HKB we are generating (i.e.,

LR A vs N R A). For both background theories we

generate a constant interval for a given variable ID

with probability 0.5 (e.g., 345 < X

3

< 789 for vari-

able ID 3). Otherwise, we pick two random subsets

of all other real variables X

L

,X

U

⊂ VARS

Real

for the

upper and lower bound respectively. Now if we are

generating an HKB with respect to the background

theory LR A, we sum all variables in the upper as

well as the lower bound, to create a linear function

as the upper and lower bound for the variable. Sim-

ilarly, when generating an HKB with respect to the

background theory N R A, we conjoin the variables

of a given set (X

L

,X

U

) by multiplication rather than

by addition. Finally, when creating such an interval

we additionally add a constant interval for the same

variable ID to make sure our integration is deﬁnite

and evaluates to a real number.

In order to evaluate our framework, we let the

number of variables (nbVars) range from 2 to 28,

where the number of clauses we tested is nbVars∗0.7,

nbVars and nbVars ∗ 1.5 for a given value of nbVars.

3

https://www.math.ucdavis.edu/

∼

latte/

4

We were unable to compare the performance with the

framework developed in (Kolb et al., 2018) owing to com-

patibility issues in the experimental setup. Since it is re-

ported to perform comparably to (Morettin et al., 2017),

all comparisons made in this paper are in reference to the

pipeline developed in (Morettin et al., 2017).

Figure 2: Runtime analysis of WMI-SDD framework for

non-linear HKBs.

Now for each variable clause pair, we generate two

problem instances where the percentage of continu-

ous variables is set to 50% to account for the random-

ness of the generator. Thus for each number of vari-

ables, we generate six different problems, which are

then averaged to compute a ﬁnal runtime.

5 RESULTS

First, we discuss the performance of our framework

on non-linear hybrid domains. As part of this experi-

ment the generated HKB consists of non-linear atoms

which are products of variables (e.g. X

1

∗ X

2

∗ −4 ∗

X

3

< X

4

< X

1

∗ 27 ∗ X

5

). Figure 2 plots the average

time spent in each computational step for all prob-

lems that have the same number of variables. Here we

see that the overall time increases with the number of

variables as expected. While most of the steps have a

rather small impact on the overall computational time,

the integration step has by far the greatest cost. This

is in part due to the Scipy integration method, which

was used for these benchmarks, as it can cope with

non-linear bounds but is not as efﬁcient as the latte

integration package. Finally, we want to point out the

surprisingly small cost of compiling the PKB into an

SDD, which reinforces our decision to use knowledge

compilation.

Next, we discuss the performance of the WMI-

SDD framework on linear HKBs against one of the

current state-of-the-art WMI solver, the WMI-PA

framework (Morettin et al., 2017). The results are

plotted in Figure 3. The results demonstrate the over-

all impact of using knowledge compilation as part of

the framework. While the additional step of compil-

ing the abstracted PKB into an SDD results in longer

computational time for small problem instances, the

trade-off shows its advantage as we increase the num-

ber of variables. Considering the logarithmic scale of

Scaling up Probabilistic Inference in Linear and Non-linear Hybrid Domains by Leveraging Knowledge Compilation

353

Figure 3: Total runtime comparison WMI-SDD vs WMI-

PA for linear HKBs.

the y-axis, the difference between the two algorithms

becomes quite substantial as the number of variables

exceeds 20. By extension, we believe the WMI-

SDD framework shows tremendous promise for scal-

ing WMI to large domains in the future.

Before concluding this section, we remark that

readers familiar with propositional model counters

are likely to be surprised by the total variable size

being less than 50 in our experiments and other

WMI solvers (Morettin et al., 2017). Contrast this

with SDD evaluations that scale to hundreds of propo-

sitional variables (Darwiche, 2011; Choi and Dar-

wiche, 2013). The main bottleneck here is symbolic

integration, even if in isolation solvers such as latte

come with strong polynomial time bounds (Baldoni

et al., 2011). This is because integration has been per-

formed for each model, and so with n variables and

a knowledge base of the form (a

1

< X

1

< b

1

) ∨ ... ∨

(a

n

< X

n

< b

n

), where a

i

,b

j

∈ R, there are 2

n

∗n inte-

gration computations in the worst case. That is, there

are 2

n

models on abstraction, and in each model, we

will have n integration variables.

There are a number of possible ways to address

that concern. First, a general solution is to simply fo-

cus on piecewise constant potentials, in which case,

after abstraction, WMI over an HKB immediately re-

duces to a WMC task over the corresponding PKB.

Second, parallelisation can be enabled. For example,

we can decompose a CNF formula into components,

which are CNF formulas themselves, the idea being

that components do not share variables (Gomes et al.,

2009). In this case, the model count of a formula F,

written #F with n components C

1

,. .. ,C

n

would be

#C

1

∗ · ·· ∗ #C

n

. This is explored for the interval frag-

ment in (Belle et al., 2016). Third, one can keep a dic-

tionary of partial computations of the integration (that

is, cache the computed integrals), and apply these val-

ues where applicable.

While we do not explore such possibilities in this

article, we feel the ability of SDDs to scale as well as

its ability to enable parallelisation can be seen as addi-

tional justiﬁcations for our approach. We also suspect

that it should be fairly straightforward to implement

such choices given the modular way our solver is re-

alized.

6 CONCLUSION

In this paper, we introduced a novel way of perform-

ing WMI by leveraging efﬁcient predicate abstrac-

tion and knowledge compilation. Using SDDs to rep-

resent the abstracted HKBs enabled us to make full

use of the structural properties of SDD and devise an

efﬁcient algorithm for retrieving all satisfying mod-

els. The evaluations demonstrate the competitiveness

of our framework and reinforce our hypothesis that

knowledge compilation is worth considering even in

continuous domains. We were also able to deal with

a speciﬁc class of separable non-linear constraints.

In the future, we would like to better explore how

the integration bottleneck can be addressed, possibly

by caching sub-integration computations. In indepen-

dent recent efforts, (Martires et al., 2019; Kolb et al.,

2019) also investigate the use of SDDs for performing

WMI. In particular, (Kolb et al., 2019) consider a dif-

ferent type of mapping between WMI and SDDs but

do not consider non-linear domains, whereas (Mar-

tires et al., 2019) allow for standard density functions

such as Gaussians by appealing to algebraic model

counting (Kimmig et al., 2016). Performing addi-

tional comparisons and seeing how these ideas could

be incorporated in our framework might be an inter-

esting direction for the future.

ACKNOWLEDGEMENTS

Anton Fuxjaeger was supported by the Engineering

and Physical Sciences Research Council (EPSRC)

Centre for Doctoral Training in Pervasive Parallelism

(grant EP/L01503X/1) at the School of Informatics,

University of Edinburgh. Vaishak Belle was sup-

ported by a Royal Society University Research Fel-

lowship. We would also like to thank our reviewers

for their helpful suggestions.

REFERENCES

Albarghouthi, A., D’Antoni, L., Drews, S., and Nori, A.

(2017). Quantifying Program Bias. arXiv e-prints,

page arXiv:1702.05437.

ICAART 2020 - 12th International Conference on Agents and Artiﬁcial Intelligence

354

Bahar, R. I., Frohm, E. A., Gaona, C. M., Hachtel, G. D.,

Macii, E., Pardo, A., and Somenzi, F. (1997). Alge-

bric decision diagrams and their applications. Formal

methods in system design, 10(2-3):171–206.

Baldoni, V., Berline, N., De Loera, J., K

¨

oppe, M., and

Vergne, M. (2011). How to integrate a polyno-

mial over a simplex. Mathematics of Computation,

80(273):297–325.

Barrett, C. W., Sebastiani, R., Seshia, S. A., and Tinelli, C.

(2009). Satisﬁability modulo theories. In (Biere et al.,

2009), pages 825–885.

Bekker, J., Davis, J., Choi, A., Darwiche, A., and Van den

Broeck, G. (2015). Tractable learning for complex

probability queries. In Advances in Neural Informa-

tion Processing Systems, pages 2242–2250.

Belle, V., Passerini, A., and Van den Broeck, G. (2015).

Probabilistic inference in hybrid domains by weighted

model integration. In Proceedings of 24th Interna-

tional Joint Conference on Artiﬁcial Intelligence (IJ-

CAI), pages 2770–2776.

Belle, V., Van den Broeck, G., and Passerini, A. (2016).

Component caching in hybrid domains with piecewise

polynomial densities. In Proceedings of the 30th Con-

ference on Artiﬁcial Intelligence (AAAI).

Biere, A., Biere, A., Heule, M., van Maaren, H., and Walsh,

T. (2009). Handbook of Satisﬁability: Volume 185

Frontiers in Artiﬁcial Intelligence and Applications.

IOS Press, Amsterdam, The Netherlands, The Nether-

lands.

Chavira, M. and Darwiche, A. (2008). On probabilistic in-

ference by weighted model counting. Artiﬁcial Intel-

ligence, 172(6-7):772–799.

Chistikov, D., Dimitrova, R., and Majumdar, R. (2017). Ap-

proximate counting in smt and value estimation for

probabilistic programs. Acta Informatica, 54(8):729–

764.

Choi, A. and Darwiche, A. (2013). Dynamic minimization

of sentential decision diagrams. In AAAI.

Choi, A., Kisa, D., and Darwiche, A. (2013). Compiling

probabilistic graphical models using sentential deci-

sion diagrams. In European Conference on Symbolic

and Quantitative Approaches to Reasoning and Un-

certainty, pages 121–132. Springer.

Darwiche, A. (2004). New advances in compiling cnf to

decomposable negation normal form. In Proceedings

of the 16th European Conference on Artiﬁcial Intelli-

gence, pages 318–322. Citeseer.

Darwiche, A. (2011). Sdd: A new canonical represen-

tation of propositional knowledge bases. In IJCAI

Proceedings-International Joint Conference on Arti-

ﬁcial Intelligence, volume 22, page 819.

Darwiche, A. and Marquis, P. (2002). A knowledge compi-

lation map. Journal of Artiﬁcial Intelligence Research,

17(1):229–264.

De Loera, J., Dutra, B., Koeppe, M., Moreinis, S., Pinto,

G., and Wu, J. (2011). Software for exact integra-

tion of polynomials over polyhedra. arXiv preprint

arXiv:1108.0117.

De Loera, J. A., Hemmecke, R., Tauzer, J., and Yoshida,

R. (2004). Effective lattice point counting in rational

convex polytopes. Journal of symbolic computation,

38(4):1273–1302.

Fierens, D., Van den Broeck, G., Renkens, J., Shterionov,

D., Gutmann, B., Thon, I., Janssens, G., and De Raedt,

L. (2015). Inference and learning in probabilistic logic

programs using weighted boolean formulas. Theory

and Practice of Logic Programming, 15(3):358–401.

Gomes, C. P., Sabharwal, A., and Selman, B. (2009). Model

counting. In (Biere et al., 2009), pages 633–654.

Kimmig, A., Van den Broeck, G., and De Raedt, L. (2016).

Algebraic model counting. International Journal of

Applied Logic.

Kisa, D., Van den Broeck, G., Choi, A., and Darwiche, A.

(2014). Probabilistic sentential decision diagrams. In

KR.

Kolb, S., Mladenov, M., Sanner, S., Belle, V., and Kersting,

K. (2018). Efﬁcient symbolic integration for proba-

bilistic inference. In IJCAI, pages 5031–5037.

Kolb, S., Zuidberg Dos Martires, P. M., and De Raedt,

L. (2019). How to exploit structure while solving

weighted model integration problems. UAI 2019 Pro-

ceedings.

Liang, Y., Bekker, J., and Van den Broeck, G. (2017).

Learning the structure of probabilistic sentential deci-

sion diagrams. In Proceedings of the 33rd Conference

on Uncertainty in Artiﬁcial Intelligence (UAI).

Martires, P., Dries, A., and De Raedt, L. (2019). Exact and

approximate weighted model integration with proba-

bility density functions using knowledge compilation.

Proceedings of the AAAI Conference on Artiﬁcial In-

telligence, 33:7825–7833.

Morettin, P., Passerini, A., and Sebastiani, R. (2017). Efﬁ-

cient weighted model integration via smt-based predi-

cate abstraction. In Proceedings of the Twenty-Sixth

International Joint Conference on Artiﬁcial Intelli-

gence, IJCAI-17, pages 720–728.

Muise, C., McIlraith, S. A., Beck, J. C., and Hsu, E. I.

(2012). D sharp: fast d-dnnf compilation with sharp-

sat. In Canadian Conference on Artiﬁcial Intelligence,

pages 356–361. Springer.

Poon, H. and Domingos, P. (2011). Sum-product networks:

A new deep architecture. In Computer Vision Work-

shops (ICCV Workshops), 2011 IEEE International

Conference on, pages 689–690. IEEE.

Sang, T., Beame, P., and Kautz, H. A. (2005). Performing

bayesian inference by weighted model counting. In

AAAI, volume 5, pages 475–481.

Sanner, S., Delgado, K., and Barros, L. (2012). Sym-

bolic dynamic programming for discrete and contin-

uous state mdps. CoRR, abs/1202.3762.

Shenoy, P. P. and West, J. C. (2011). Inference in hy-

brid bayesian networks using mixtures of polynomi-

als. International Journal of Approximate Reasoning,

52(5):641–657.

Suciu, D., Olteanu, D., R

´

e, C., and Koch, C. (2011). Proba-

bilistic databases. Synthesis lectures on data manage-

ment, 3(2):1–180.

Van den Broeck, G. and Darwiche, A. (2015). On the role of

canonicity in knowledge compilation. In AAAI, pages

1641–1648.

355