Transcript Dynamic Restart Policy

Dynamic Restarts
Optimal Randomized Restart Policies
with Observation
Henry Kautz, Eric Horvitz, Yongshao Ruan,
Carla Gomes and Bart Selman
Outline

Background



Optimal strategies to improve expected time
to solution using



heavy-tailed run-time distributions of backtracking
search
restart policies
observation of solver behavior during particular
runs
predictive model of solver performance
Empirical results
Backtracking Search

Backtracking search algorithms often
exhibit a remarkable variability in
performance among:




slightly different problem instances
slightly different heuristics
different runs of randomized heuristics
Problematic for practical application

Verification, scheduling, planning
Heavy-tailed Runtime Distributions

Observation (Gomes 1997): distributions of runtimes
of backtrack solvers often have heavy tails


infinite mean and variance
probability of long runs decays by power law (ParetoLevy), rather than exponentially (Normal)
Very short
Very long
Formal Models of Heavy-tailed
Behavior

Imbalanced tree search models (Chen 2001)


Exponentially growing subtrees occur with
exponentially decreasing probabilities
Heavy-tailed runtime distribution can arise in
backtrack search for imbalanced models with
appropriate parameters p and b


p is the probability of the branching heuristics
making an error
b is the branch factor
Randomized Restarts

Solution: randomize the systematic solver




Provably eliminates heavy tails
Effective whenever search stagnates


Add noise to the heuristic branching (variable
choice) function
Cutoff and restart search after some number
of steps
Even if RTD is not formally heavy-tailed!
Used by all state-of-the-art SAT engines


Chaff, GRASP, BerkMin
Superscalar processor verification
Complete Knowledge of RTD
P(t)
D
t
Complete Knowledge of RTD
Luby (1993): Optimal policy uses fixed cutoff
T *  arg min t E ( Rt )
where E ( Rt ) is the expected time to solution
restarting every t steps
P(t)
D
T*
t
Complete Knowledge of RTD
Luby (1993): Optimal policy uses fixed cutoff
T *  arg min E (Tt )
t
t
E (Tt ) 
 t  E (T | T  t )  where T is the
P(T  t )
length of single complete run (without cutoff)
P(t)
D
T*
t
No Knowledge of RTD
Luby (1993): Universal sequence of cutoffs
1, 1, 2, 1, 1, 2, 4, ...
is within O(log T *) of the optimal policy for the
unknown distribution
In practice - 1-2 orders of magnitude

Open cases:


Partial knowledge of RTD (CP 2002)
Additional knowledge beyond RTD
Example: Runtime
Observations

Idea: use observations of early
progress of a run to induce finergrained RTD’s
D1
P(t)
T1
D
T*
D2
T2
t
Example: Runtime
Observations
What is optimal policy, given original & component
RTD’s, and classification of each run?
 Lazy: use static optimal cutoff for combined RTD
P(t)
D1
D
T*
D2
t
Example: Runtime
Observations
What is optimal policy, given original & component
RTD’s, and classification of each run?
 Naïve: use static optimal cutoff for each RTD
D1
P(t)
T1*
D2
T2*
t
Results

Method for inducing component distributions
using





Bayesian learning on traces of solver
Resampling & Runtime Observations
Optimal policy where observation assigns each
run to a component distribution
Conditions under which optimal policy prunes
one (or more) distributions
Empirical demonstration of speedup
I. Learning to Predict
Solver Performance
Formulation of Learning
Problem

Consider a burst of evidence over
observation horizon

Learn a runtime predictive model
using supervised learning
Observation horizon
Short
Long
Median run time
Horvitz, et al. UAI 2001
Runtime Features

Solver instrumented to record at each choice
(branch) point:



SAT & CSP generic features: number free
variables, depth of tree, amount unit propagation,
number backtracks, …
CSP domain-specific features (QCP): degree of
balance of uncolored squares, …
Gather statistics over 10 choice points:



initial / final / average values
1st and 2nd derivatives
SAT: 127 variables, CSP: 135 variables
Learning a Predictive Model

Training data: samples from original
RTD labeled by
(summary features, length of run)


Learn a decision tree that predicts
whether current run will complete in
less than the median run time
65% - 90% accuracy
Generating Distributions by
Resampling the Training Data

Reasons:



The predictive models are imperfect
Analyses that include a layer of error analysis for
the imperfect model are cumbersome
Resampling the training data:


Use the inferred decision trees to define different
classes
Relabel the training data according to these
classes
Creating Labels


The decision tree reduces all the observed
features to a single evidential feature F
F can be:

Binary valued


Indicates prediction: shorter than median runtime?
Multi-valued

Indicates particular leaf of the decision tree that is
reached when trace of a partial run is classified
Result

P(t)
Decision tree can be used to precisely
classify new runs as random samples
from the induced RTD’s
Observed F
Make Observation
D
median
Observed F
t
II. Creating Optimal Control
Policies
Control Policies

Problem Statement:





A process generates runs randomly from a
known RTD
After the run has completed K steps, we may
observe features of the run
We may stop a run at any point
Goal: Minimize expected time to solution
Note: using induced component RTD’s
implies that runs are statistically
independent

Optimal policy is stationary
Optimal Policies
Optimal restart policy for using a binary
feature F is of the form:
(1) Set cutoff to T1 for a fixed T1  TObs
or of the form:
(2) Wait for TObs steps, then observe F;
If F holds, then use cutoff T1
else use cutoff T2
for appropriate constants T1 , T2
Straightforward generalization to multi-valued features
Case (2): Determining Optimal
Cutoffs
(T1** , T2** ,...)  arg min E (T1 , T2 ,...)
Ti
 d (T   q (t ))
i
 arg min
Ti
i
i
i
t Ti
 d q (T )
i i
i
i
where d i is the probability of observing the i -th value of F
q (t ) is the probability run succeeds in t or fewer steps
Optimal Pruning

Runs from component D2 should be pruned
(terminated) immediately after observation
when:
for all   0 : E (T1 , TObs   )  E (T1 , TObs )
Equivalently:
 - T
Obs  t TObs
q
(
t
)
2

q2 (TObs  )  q2 (TObs )
 E (T1 , TObs )
Note: does not depend on priors on Di
III. Empirical Evaluation
Backtracking Problem Solvers

Randomized SAT solver

Satz-Rand, a randomized version of Satz (Li 1997)

DPLL with 1-step lookahead


Randomization with noise parameter for increasing
variable choices
Randomized CSP solver

Specialized CSP solver for QCP

ILOG constraint programming library

Variable choice, variant of Brelaz heuristic
Domains

Quasigroup With Holes

Graph Coloring

Logistics Planning (SATPLAN)
Dynamic Restart Policies

Binary dynamic policies


Runs are classified as either having
short or long run-time distributions
N-ary dynamic policies

Each leaf in the decision tree is
considered as defining a distinct
distribution
Policies for Comparison

Luby optimal fixed cutoff



For original combined distribution
Luby universal policy
Binary naïve policy

Select distinct, separately optimal fixed
cutoffs for the long and for the short
distributions
Illustration of Cutoffs
D1
P(t)
T1*
T2**
Make Observation
D
T*
T1**
D2
T2*
t
Comparative Results
Improvement of dynamic policies over Luby fixed optimal
cutoff policy is 40~65%
Expected Runtime (Choice Points)
QCP (CSP)
QCP (Satz)
Graph Coloring (Satz)
Planning (Satz)
Dynamic n-ary
3,295
8,962
9,499
5,099
Dynamic binary
5,220
11,959
10,157
5,366
Fixed optimal
6,534
12,551
14,669
6,402
Binary naïve
17,617
12,055
14,669
6,962
Universal
12,804
29,320
38,623
17,359
Median (no cutoff)
69,046
48,244
39,598
25,255
Cutoffs: Graph Coloring (Satz)
Dynamic n-ary:
Dynamic binary:
Binary naive:
Fixed optimal:
10, 430, 10, 345, 10, 10
455, 10
342, 500
363
Discussion


Most optimal policies turned out to
prune runs
Policy construction independent from
run classification – may use other
learning techniques


Does not require highly-accurate
prediction!
Widely applicable
Limitations


Analysis does not apply in cases where runs
are statistically dependent
Example:

We begin with 2 or more RTD’s


Environment flips a coin to choose a RTD, and
then always samples that RTD


E.g.: of SAT and UNSAT formulas
We do not get to see the coin flip!
Now each unsuccessful run gives us information
about that coin flip!
The Dependent Case


Dependent case much harder to solve
Ruan et al. CP-2002:


“Restart Policies with Dependence among
Runs: A Dynamic Programming Approach”
Future work


Using RTD’s of ensembles to reason about
RTD’s of individual problem instances
Learning RTD’s on the fly (reinforcement
learning)
Big Picture
control / policy
runtime
Problem
Instances
Solver
dynamic
features
static features
Learning /
Analysis
Predictive
Model