Slides - Stanford Computer Science

Download Report

Transcript Slides - Stanford Computer Science

Computing the
Density of States of
Boolean Formulas
Stefano Ermon,
Carla Gomes, and Bart Selman
Cornell University, September 2010
Motivation:
Significant progress in SAT
From 100 variables, 200 constraints (early 90’s)
to over 1,000,000 vars. and 5,000,000 clauses
in 20 years.
Applications:
Hardware and Software Verification, Planning,
Scheduling, Optimal Control, Protocol Design,
Routing, Multi-agent systems, E-Commerce
(E-auctions and electronic trading agents), etc.
SAT: Given a Boolean formula Φ in CNF,
Φ=C1ΛC2 Λ…ΛCm
does Φ have a satisfying assignment?
Extending SAT technology

Model counting problem (number of distinct
satisfying assignments):
probabilistic inference problems
 multi-agent / adversarial reasoning (bounded)
[Roth ‘96, Littman et. al. ‘01, Sang et. al. ‘04, Darwiche ‘05,
Domingos ‘06]


MAX-SAT and Weighted MAX-SAT: find a truth
assignment that maximizes the number of
satisfied clauses or the sum of their weights


beyond decision (NP) [Hansen at al. ’90]
hard and soft constraints [Heras et al. ’08, Cohen et al. ’06]
How can we combine both challenges?
Density of states

Given a Boolean formula Φ in CNF,
Φ=C1ΛC2 Λ…ΛCm with m clauses
the density of states is a function
n : [0,..,m]  
that gives the number of truth assignments that
violate exactly i clauses, for i =0,..,m


n(0) = number of assignments that violate 0 clauses (models)
n(1) = number of assignments that violate exactly 1 clause
Density of states: a challenging
problem

Generalizes SAT


Generalizes MAX-SAT


Decision problem: Φ is satisfiable if and only if n(0)>0
MAX-SAT is the minimum i such that n(i)>0
Generalizes #SAT

Number of models = n(0)
Statistical physics

More generally, the density of states (DOS)
gives the number of microstates with energy E
 Microstates
= truth assignments
 Energy = number of violated clauses
 Ground states = maximally satisfying assignments

Compact, very informative characterization of a
physical system
 Macroscopic
thermodynamic quantities (free energy,
internal energy,..)
 Partition function, phase transitions,..
Motivation


DOS provides a finer characterization of the
structure of a combinatorial search space
Statistical physics and CSPs:
 Insights
on problem structure, hardness, new
algorithms, Survey Propagation [Montanari et. al. ‘07,
Monasson et. al. ‘96, Mézard ’02, Parisi ‘02]

By defining different energy functions, it can be
naturally used for probabilistic style inference
(e.g. Markov Logic, [Domingos ‘06] )
Talk Outline



Prior work
A novel sampling strategy: MCMC-FlatSat
Empirical Validation
 Small
formulas with ground truth
 Synthetic formulas
 Random 3-SAT
 Large structured instances
 Model counting

Conclusions
Density of states: prior work


Exact Method: Enumeration (exponential)
Approximate
 Uniform


Sampling [Belaidouni et. al. ‘02]
Sample density from K random truth assignments
Impractical, unlikely to hit rare assignments (e.g. solutions)
 Metropolis


Sampling [Rose et. al. ‘96]
In theory, the density can be extracted from the Boltzmann
distribution
Impractical, difficult choice of the temperature and slow
mixing times [Wei et. al. ‘04 ]
How can we improve the sampling strategy?
The flat histogram idea

Idea: Set up a Markov Chain that visits all
energy levels equally often [F. Wang and Landau
’02, J. Wang et. al. ’99, De Oliveira et. al. ’96]
e.g. an equal amount of time at the set of truth
assignments with 0 unsat clauses, 1 unsat clause, ...


How? Flip a variable, accept new state σ’ with
probability
 n( E ) 
p    min1,

 n( E) 
Always accepts “rarer” states (when n(E’)<n(E))
Example
RED
↔ 0 unsat clauses, n(0)=1
GREEN ↔ 1 unsat clauses, n(1)=3
BLUE ↔ 2 unsat clauses, n(2)=10
1
1
1
1
1
Goal: visit all energy levels
(colors) equally often
Example
RED
↔ 0 unsat clauses, n(0)=1
GREEN ↔ 1 unsat clauses, n(1)=3
BLUE ↔ 2 unsat clauses, n(2)=10
3/10
3/10
3/10
3/10
Goal: visit all energy levels
(colors) equally often
Example
RED
↔ 0 unsat clauses, n(0)=1
GREEN ↔ 1 unsat clauses, n(1)=3
BLUE ↔ 2 unsat clauses, n(2)=10
3/10
3/10
3/10
3/10
Goal: visit all energy levels
(colors) equally often
Example
RED
↔ 0 unsat clauses, n(0)=1
GREEN ↔ 1 unsat clauses, n(1)=3
BLUE ↔ 2 unsat clauses, n(2)=10
1
1
1
Goal: visit all energy levels
(colors) equally often
Example
RED
↔ 0 unsat clauses, n(0)=1
GREEN ↔ 1 unsat clauses, n(1)=3
BLUE ↔ 2 unsat clauses, n(2)=10
1
1
1
1
1
Goal: visit all energy levels
(colors) equally often
Example
RED
↔ 0 unsat clauses, n(0)=1
GREEN ↔ 1 unsat clauses, n(1)=3
BLUE ↔ 2 unsat clauses, n(2)=10
1
1
1
1
1
Goal: visit all energy levels
(colors) equally often
Example
RED
↔ 0 unsat clauses, n(0)=1
GREEN ↔ 1 unsat clauses, n(1)=3
BLUE ↔ 2 unsat clauses, n(2)=10
1
1
1
1
1
Goal: visit all energy levels
(colors) equally often
Example
RED
↔ 0 unsat clauses, n(0)=1
GREEN ↔ 1 unsat clauses, n(1)=3
BLUE ↔ 2 unsat clauses, n(2)=10
1/10
1/3
1/10
1/10
1
Goal: visit all energy levels
(colors) equally often
Example
RED
↔ 0 unsat clauses, n(0)=1
GREEN ↔ 1 unsat clauses, n(1)=3
BLUE ↔ 2 unsat clauses, n(2)=10
1/10
1/3
1/10
1/10
1
Goal: visit all energy levels
(colors) equally often
Example
…and finally…
Example
RED
↔ 0 unsat clauses, n(0)=1
GREEN ↔ 1 unsat clauses, n(1)=3
BLUE ↔ 2 unsat clauses, n(2)=10
Flat histogram: Intuition



Detailed balance holds with those
 n( E ) 
transition probabilities
p    min1,

 n( E) 
Properly biased towards “rare” states
But there are few “rare” states
(Red, Blue, Green)
The total time spent in each type of state is
the same (flat visit histogram).
Note: in contrast, Simulated Annealing concentrates sampling
around low energy states (more greedy!)
Flat histogram
But… how can we even run the Markov Chain?
Acceptance probability:
p  
 n( E ) 
 min1,

 n( E) 
The density n() is unknown and is precisely what
we want to compute!
Adaptive sampling


Start with an initial guess g (our estimate of the
true density n)
Random walk:
 Use
g for guidance (acceptance probability)
 Chain will initially not sample uniformly across energy
levels
 Each step, adjust g using a modification factor F
 Keep track of the visit histogram H

When we see a flat H, we must have the right
density!
The modification factor, F


F controls the tradeoff between convergence
rate and accuracy
Use large modification factors F at the beginning
to get
 rough
estimates
 fast convergence


Keep reducing F to get finer estimates
Analogous to an annealing process
MCMC-FlatSat
Initialization
Inner
Loop
H
100
96
105
99
50
0
0 (R ed)
1
2 (B lue)
(G reen)
Inner Loop : adaptive sampling until the visit histogram is flat
(g becomes our new guess for n)
MCMC-FlatSat
Outer
Loop
Reduce modification factor and repeat inner loop until g≈n
Outline of empirical validation

Empirical validation on combinatorial problems
 Convergence
 Efficiency (number
 Accuracy

of samples vs search space size)
We study:
 Small formulas with ground
 Large synthetic formulas
 Random 3-SAT
 Large structured
 Model counting
instances
truth
Formulas with known ground truth



Instances from MAXSAT2007 competition
(Ramsey, Spin Glass, Max Clique)
Direct enumeration is possible (n<=28), so we
can compare our estimate with ground truth
Metrics:
n( E )  n( E ) 
 KL divergence: DKL (n || g )  

log
Z
E
 g (E) 
 Relative
error per point
n – ground truth
X g – estimate
Rel. error (%)
Log-density
3.1 % maximum relative error
Energy
Energy (# unsat clauses)


Spin glass instance, 27 variables, 162 clauses
Needs ~ 8 *106 flips << search space size
(227≈1.3 *108)
Max Clique
Ramsey
n
X g
Log-density
Log-density
n – ground truth
X g – estimate
Energy (# unsat clauses)
Energy (# unsat clauses)
Instance
variables
clauses
KLdivergence
ramk3n7.ra0
21
70
3.9 E-05
2.39 %
2.45
ramk3n8.ra0
28
126
1.1 E-05
5.1 %
3.93
johnson8-2-4.clq
28
420
4.5 E-05
5.5 %
2.90
T3pm3-5555.spn
27
162
1.3 E-05
3.15 %
3.34
Max rel. error
Entropy
Synthetic formulas

Ground truth for larger formulas?
 Construct



synthetic formulas
Formulas for which we derive a closed form
solution for the density
Result on the composition (logical conjunction)
of independent formulas (do not share variables)
F ( x1, x2 ,...,xl )   ( x1 )   ( x2 )  ...  ( xl )
The density of F is the convolution of the
density of Φ with itself l times
Convolution of
uniform densities
n
X g
Log-density
Log-density
n – ground truth
X g – estimate
Convolution of
pigeon hole
formulas
Energy (# unsat clauses)
Energy (# unsat clauses)
Needs ~ 2*107 flips << state space size (250≈1015)
Instance
variables
clauses
KL-divergence Max relative error (%)
Entropy
UnifConv50_100
50
100
1.19 E-05
3.05 %
7.3
PigHoleConv410
200
750
1.26 E-07
2.2 %
33.7
Random k-SAT formulas


Well known phase transitions for the
satisfiability property P [n (0)  0] in terms of the
ratio α (Φ is satisfiable ↔ nΦ(0)>0 )
Analytic result on the average density
 m  1  
1
E [n (i)]    k  1  k 
 i  2   2 
i
m i
2n
given a truth assignment, the probability of
having a clause that is violated is 1/2k for
random k-SAT
Log-density
Average Densities
Ratio clauses to var. α


n=50 variables (average over 1000 instances)
Needs ~ 108 flips << state space size (250≈1015)
Large structured instances


No ground truth known
Consistency checks
 Number
of models, when exact model counting is
feasible (# models=n(0))
 Method of the moments:
 Sample K assignments at random, compute their
energies (unsat clauses)
 Compute sample moments (e.g. average energy,
2nd order moment of energy, ..)
 Compare with moments obtained using the
estimated density g
Large structured instances
Instance
variables clauses g(0)
#
Ms(1)
models
M(1)
Ms(2)
M(2)
brock400 2.clq
40
1188
0
0
297.0
297.0
88365.9
88372.3
Spinglass5.pm
125
750
0
0
187.4
187.4
35249.2
35247.1
MANN a27.clq
42
1690
0
0
422.4
422.4
178709
178703
bw large.a
459
4675
1
1
995.2
995.3
996349
996634

Planning problem:
 n=459 variables
 m=4675 clauses
Log-density
Logistic instance
Energy (# unsat clauses)


Huge search space (2459 truth assignments), but
MCMC-FlatSat returns within hours
Remarkable precision:
 Finds
the only existing model g(0)=n(0)=1!
 The mode of the estimated distribution is e300 times
larger than the number of models g(0)
 counts the needles and the haystack!
(the moments method indicates g is accurate)
Model counting

Comparison with state-of-the-art model counters
[Gomes et. al. ‘06]
 SampleMiniSAT [Gogate et. al. ‘07]
 SampleCount


MCMCFlatSat is very accurate
Timings are competitive when ratio clauses to
variables is not too large
 DOS
provides guidance
 Information on what is not a model
 Overhead because it provides more information
Model counting comparison
SampleCount SampleMiniSAT MCMC-FlatSat
Instance variables clauses
Exact #
models
Models
Time
(s)
Models
Time
(s)
Models
Time
(s)
2bitmax
252
766
2.10×1029 >2.40×1028
29
2.08×1029
345 1.96×1029 1863
wff-3-3.5
150
525
1.40×1014 >1.60×1013 145
1.60×1013
240 1.34×1014
393
wff-3.1.5
100
150
1.80×1021 >1.00×1020 240
1.58×1021
128 1.83×1021
21
wff-4-5.0
100
500
>8.00×1015 120
1.09×1017
191 8.64×1016
189
ls8-norm
301
1603
5.40×1011 >3.10×1010 1140
2.22×1011
168
5.93×1011 2693
Conclusions


Computing the density of states is a hard
problem (encompasses SAT, MAX-SAT, #SAT)
MCMCFlatSat: sampling strategy adapted from
physics for combinatorial spaces that adaptively
explores the space while collecting statistics
 Extremely
accurate, very efficient (few samples)
 Provides a compact, rich description of the search
space. New insights about structure and local search

Very general method: any property can be
used for search space partitioning.
 Many
applications to counting and inference problems
Extra slides
Future work

SAT-specific improvements
 Energy
saturation
 Energy barriers (and related normalization issues)
 Walksat heuristics (with Metropolis-Hastings updates)



Direct application to inference in Markov Logic
Formal proof of convergence / counterexamples
Application to other counting problems in
combinatorial spaces
Runtime for random 3-SAT


Search space 2^50
Flips 10^8 ≈ 2^26
Related work on random 3-SAT

Lots of work on the i=0 (i.e. SAT/UNSAT) case
[Gent et. al. ‘94]

Previous experimental work for i>0 [Zhang ‘01]
 Different
definition: “no more than i unsat clauses”
versus “exactly i unsat clauses”
 Location on the phase transitions
 Apparently, same location
 Analytic results [Achlioptas et al. ‘05]

We can see two phase transitions: assignments
are unlikely to violate a large number of clauses
Other phase transition
50 variables, 1000 instances
Other structured instances
Histogram flatness (formal)

Flatness condition of the visit histogram H is a
necessary condition for convergence
 If
g is equal to the true density n, the detailed balance
p( ) p    p( ) p 
1
is satisfied by p( ) 
n( E ( ))
upon convergence, the steady state probability is
proportional to the reciprocal of the density of the
corresponding energy level
Histogram flatness (formal)


Flatness condition of the visit histogram H is a
necessary condition for convergence
c
H ( E )   p( )  
c
 | E ( )  E
 | E ( )  E n( E )
If g=n, the visit histogram H will be flat
i.e. energy levels visited equally often
Problem: the density n() is unknown and is
precisely what we want to compute!
Propositional Satisfiability (SAT)
Satifiability (SAT): Given a formula in propositional calculus,
does it have a model, i.e., is there an assignment to its
variables making it true?
(a

b  c) AND (b  c) AND (a  c)
SAT: prototypical hard combinatorial search and reasoning
problem. Problem is NP-Complete. (Cook 1971)
Prior work - Metropolis

Approximate [Rose et. al., ‘96 ]
 Metropolis
sampling at temperature T
 Energy = number of violated clauses
 Measure the empirical probability P(E) of energy
levels, rescale to get the density n(E)
P( E )  n( E ) exp( E )
T
mixing times [Wei et. al., ‘04 ]
 Difficult choice of the temperature
 Slow
Theoretical Convergence

Stated as a conjecture in
Lee,Okabe and Landau, Convergence and
Refinement of the Wang-Landau Algorithm,
Computer Physics Communications, 2006

Proof of convergence in
Atchade and Liu,The Wang-Landau Algorithm
for Monte Carlo computation in general state
spaces, Statistica Sinica, 2009
Closed forms

k-SAT, m clauses such that each variable
appears exactly in one clause, the DOS is
 m  1  
1
n( E )    k  1  k 
 E  2   2 
E
m E
2n
where 1 2 is the fraction of assignments of the
variables in a single clause not satisfying it (e.g.
in 3-SAT, only 1 over 8 assignments does not
satisfy a clause)
k
Implementation details




Use of log-densities
New assignment generated by flipping a variable
uniformly at random
Flatness condition (within 10% of the max)
Normalization
 The
density is obtained only up to a constant factor
g ( E )  2n
 We use the normalization constraint



E
F0=1.5, reduce by F  F
Flatness is checked every 1000 moves
Random k-SAT formulas


k-SAT: each clause has at most k literals
Random k-SAT: formulas Φ generated with
n
variables
 m clauses produced independently by


α

randomly choosing a set of k variables from n available
negating each one with probability 0.5
is the ratio clauses to variables (m/n)
Let P () be the probability measure associated
with this generative model for Φ
Composition

Logical conjunction of formulas Φ that do not
share variables
F ( x1, x2 ,...,xl )   ( x1 )   ( x2 )  ...  ( xl )

The density of F is the convolution of the
density of Φ with itself l times
Analogous to the probability density of the sum
of l independent random variables

Closed forms

Starting with a formula Φ with uniform density
(e.g.  ( x1, x2 )  x1  x2  ( x1  x2 )  ( x1  x2 ) )
we construct
F ( x1 , x2 ,...,xl )   ( x1 , x2 )   ( x1, x2 )  ...  ( xl 1, xl )
Sum of n s-sided dices has closed form probability
distribution: