Probabilistic Modeling for Combinatorial Optimization

1
Probabilistic Modeling for Combinatorial
Optimization

• Scott Davies
• School of Computer Science
• Carnegie Mellon University
• Joint work with Shumeet Baluja

2
Combinatorial Optimization

• Maximize evaluation function f(x)
• input fixed-length bitstring x
• output real value
• x might represent
• job shop schedules
• TSP tours
• discretized numeric values
• etc.
• Our focus Black Box optimization
• No domain-dependent heuristics

x1001001...
f(x)
37.4
3
Most Commonly Used Approaches

• Hill-climbing, simulated annealing
• Generate candidate solutions neighboring single
  current working solution (e.g. differing by one
  bit)
• Typically make no attempt to model how particular
  bits affect solution quality
• Genetic algorithms
• Attempt to implicitly capture dependency of
  solution quality on bit values by maintaining a
  population of candidate solutions
• Use crossover and mutation operators on
  population members to generate new candidate
  solutions

4
Using Explicit Probabilistic Models

• Maintain an explicit probability distribution P
  from which we generate new candidate solutions
• Initialize P to uniform distribution
• Until termination criteria met
• Stochastically generate K candidate solutions
  from P
• Evaluate them
• Update P to make it more likely to generate
  solutions similar to the good solutions
• Several different choices for what sorts of P to
  use and how to update it after candidate solution
  evaluation

5
Probability Distributions Over Bitstrings

• Let x (x1, x2, , xn), where xi can take one of
  the values 0, 1 and n is the length of the
  bitstring.
• Can factorize any distribution P(x1xn) bit by
  bit
• P(x1,,xn) P(x1) P(x2 x1) P(x3 x1,
  x2)P(xnx1, , xn-1)
• In general, the above formula is just another way
  of representing a big lookup table with one entry
  for each of the 2n possible bitstrings.
• Obviously too many parameters to estimate from
  limited data!

6
Representing Independencies withBayesian Networks

• Graphical representation of probability
  distributions
• Each variable is a vertex
• Each variables probability distribution is
  conditioned only on its parents in the directed
  acyclic graph (dag)

Wean on Fire
P(F,D,I,A,H) P(F) P(D) P(IF) P(AF)
P(HD,I,A)
F
Ice Cream Truck Nearby
Fire Alarm Activated
D
I
A
Office Door Open
H
Hear Bells
7
Bayesian Networks for Bitstring Optimization?

• P(x1,,xn) P(x1) P(x2 x1) P(x3 x1,
  x2)P(xnx1, , xn-1)

Yuck. Lets just assume all the bits are
independent instead. (For now.)
P(x1,,xn) P(x1) P(x2) P(x3)P(xn)
Ahhh. Much better.
8
Population-Based Incremental Learning

• Population-Based Incremental Learning (PBIL)
  Baluja, 1995
• Maintains a vector of probabilities one
  independent probability P(xi) for each bit xi.
• Until termination criteria met
• Generate a population of K bitstrings from P
• Evaluate them
• Use the best M of the K to update P as follows

if xi is set to 1
if xi is set to 0
or

• Optionally, also update P similarly with the
  bitwise complement of the worst of the K
  bitstrings
• Return best bitstring ever evaluated

9
PBIL vs. Discrete Learning Automata

• Equivalent to a team of Discrete Learning
  Automata, one automata per bit. Thathachar
  Sastry, 1987
• Learning automata choose actions independently,
  but receive common reinforcement signal dependent
  on all their actions
• PBIL update rule equivalent to linear
  reward-inaction algorithm Hilgard Bower, 1975
  with success defined as best in the bunch
• However, Discrete Learning Automata typically
  used previously in problems with few variables
  but noisy evaluation functions

10
PBIL vs. Genetic Algorithms

• PBIL originated as tool for understanding GA
  behavior
• Similar to Bit-Based Simulated Crossover (BSC)
  Syswerda, 1993
• Regenerates P from scratch after every generation
• All K used to update P, weighted according to
  probabilities that a GA would have selected them
  for reproduction
• Why might normal GAs be better?
• Implicitly capture inter-bit dependencies with
  population
• However, because model is only implicit,
  crossover must be randomized. Also, limited
  population size often leads to premature
  convergence based on noise in samples.

11
Four Peaks Problem

• Problem used in Baluja Caruana, 1995 to test
  how well GAs maintain multiple solutions before
  converging.
• Given input vector X with N bits, and difficulty
  parameter T
• FourPeaks(T,X)MAX(head(1,X), tail(0,X))Bonus(T,X
  )
• head(b,X) of contiguous leading bits in X set
  to b
• tail(b,X) of contiguous trailing bits in X
  set to b
• Bonus(T,X) 100 if (head(1,X)gtT) AND (tail(0,X)
  gt T), or 0 otherwise
• Should theoretically be easy for GA to handle
  with single-point crossover

12
Four Peaks Problem Results

• Insert nasty photocopy here

13
Large-Scale PBIL Empirical Comparison

• See bug-ugly photocopied table

14
Modeling Inter-Bit Dependencies

• How about automatically learning probability
  distributions in which at least some dependencies
  between variables are modeled?
• Problem statement given a dataset D and a set of
  allowable Bayesian networks Bi, find the Bi
  with the maximum posterior probability

15
Equations. (Mwuh hah hah hah!)

• Where
• d j is the jth datapoint
• dij is the value assigned to xi by d j.
• Pi is the set of xis parents in B

• is the set of values assigned to Pi by d j
• P is the empirical probability distribution
  exhibited by D

16
Mmmm...Entropy.
Pick B to Maximize

• Fact given that B has a network structure S, the
  optimal probabilities to use in B are just the
  probabilities in D, i.e. P. So

Pick S to Maximize
17
Entropy Calculation Example
x20s contribution to score
80 total datapoints
where H(p,q) p log p q log q.
18
Single-Parent Tree-Shaped Networks

• Now lets allow each bit to be conditioned on at
  most one other bit.

• Adding an arc from xj to xi increases network
  score by
• H(xi) - H(xixj) (xjs information gain
  with xi)
• H(xj) - H(xjxi) (not necessarily obvious,
  but true)
• I(xi, xj) Mutual information
  between xi and xj

19
Optimal Single-Parent Tree-Shaped Networks

• To find optimal single-parent tree-shaped
  network, just find maximum spanning tree using
  I(xi, xj) as the weight for the edge between xi
  and xj. Chow and Liu, 1968
• Start with an arbitrary root node xr.
• Until all n nodes have been added to the tree
• Of all pairs of nodes xin and xout, where xin has
  already been added to the tree but xout has not,
  find the pair with the largest I(xin, xout).
• Add xout to the tree with xin as its parent.
• Can be done in O(n2) time (assuming D has already
  been reduced to sufficient statistics)

20
Optimal Dependency Trees for Combinatorial
Optimization

• Baluja Davies, 1997
• Start with a dataset D initialized from the
  uniform distribution
• Until termination criteria met
• Build optimal dependency tree T with which to
  model D.
• Generate K bitstrings from probability
  distribution represented by T. Evaluate them.
• Add best M bitstrings to D after decaying the
  weight of all datapoints already in D by a factor
  a between 0 and 1.
• Return best bitstring ever evaluated.
• Running time O(Kn n2) per iteration

21
Tree-based Optimization vs. MIMIC

• Tree-based optimization algorithm inspired by
  Mutual Information Maximization for Input
  Clustering (MIMIC) De Bonet, et al., 1997
• Learned chain-shaped networks rather than
  tree-shaped networks
• Dataset best N of all bitstrings ever evaluated
• dataset has simple, well-defined
  interpretation
• - have to remember bitstrings
• - seems to converge too quickly on some larger
  problems

x4
x8
x17
x5
22
MIMIC dataset vs. Exp. Decay dataset

• Insert skanky photocopy here

23
Graph-Coloring Example

• Photocopy or PostScript inclusion

24
Peaks problems results

• Postscript/photcopy

25
Tree-Max problem results

• Yet another dummy slide

26
Checkerboard problem results

• Dummy slide

27
Linear Equations Results

• Yucky photocopy maybe

28
Modeling Higher-Order Dependencies

• The maximum spanning tree algorithm gives us the
  optimal Bayesian Network in which each node has
  at most one parent.
• What about finding the best network in which each
  node has at most K parents for Kgt1?
• NP-complete problem! Chickering, et al., 1995
• However, can use search heuristics to look for
  good network structures (e.g., Heckerman, et
  al., 1995), e.g. hillclimbing.

29
Scoring Function for Arbitrary Networks

• Rather than restricting K directly, add penalty
  term to scoring function to limit total network
  size
• Equivalent to priors favoring simpler network
  structures
• Alternatively, lends itself nicely to MDL
  interpretation
• Size of penalty controls exploration/exploitation
  tradeoff

l penalty factor
B number of parameters in B
30
Bayesian Network-Based Combinatorial Optimization

• Initialize D with C bitstrings from uniform
  distribution, and Bayesian network B to empty
  network containing no edges
• Until termination criteria met
• Perform steepest-ascent hillclimbing from B to
  find locally optimal network B. Repeat until no
  changes increase score
• Evaluate how each possible edge addition, removal
  or deletion would affect penalized log-likelihood
  score
• Perform change that maximizes increase in score.
• Set B to B.
• Generate and evaluate K bitstrings from B.
• Decay weight of datapoints in D by a.
• Add best M of the K recently generated datapoints
  to D.
• Return best bit string ever evaluated.

31
Cutting Computational Costs

• Can cache contingency tables for all possible
  one-arc changes to network structure
• Only have to recompute scores associated with at
  most two nodes after arc added, removed, or
  reversed.
• Prevents having to slog through the entire
  dataset recomputing score changes for every
  possible arc change when dataset changes.
• However, kiss memory goodbye
• Only a few network structure changes required
  after each iteration since dataset hasnt changed
  much (one or two structural changes on average)

32
Evolution of Network Complexity

• Placeholder for parity-based network complexity
  graph

33
Summation Cancellation

• Minimize magnitudes of cumulative sum of
  discretized numeric parameters (s1, , sn)
  represented with standard binary encoding

Average value over 50 runs of best solution found
in 2000 generations
34
Bayesian Networks Empirical Results Summary

• Does better than Tree-based optimization
  algorithm on some toy problems
• Significantly better on Summation Cancellation
  problem
• 10 reduction in error on System of Linear
  Equation problems
• Roughly the same as Tree-based algorithm on
  others, e.g. small Knapsack problems
• Significantly more computation despite efficiency
  hacks, however.
• Why not much better performance?
• Too much emphasis on exploitation rather than
  exploration?
• Steepest-ascent hillclimbing over network
  structures not good enough?

35
Using Probabilistic Models for Intelligent
Restarts

• Tree-based algorithms O(n2) execution time per
  generation very expensive for large problems
• Even more so for more complicated Bayesian
  networks
• One possible approach use probabilistic models
  to select good starting points for faster
  optimization algorithms, e.g. hillclimbing

36
COMIT

• Combining Optimizers with Mutual Information
  Trees Baluja Davies, 1997b
• Initialize dataset D with bitstrings drawn from
  uniform distribution
• Until termination criteria met
• Build optimal dependency tree T with which to
  model D.
• Generate K bitstrings from the distribution
  represented by T. Evaluate them.
• Execute a hillclimbing run starting from single
  best bitstring of these K.
• Replace up to M bitstrings in D with the best
  bitstrings found during the hillclimbing run.
• Return best bitstring ever evaluated

37
COMIT, contd

• Empirical tests performed with stochastic
  hillclimbing algorithm that allows at most
  PATIENCE moves to points of equal value before
  restarting
• Compare COMIT vs.
• Hillclimbing with restarts from bitstrings chosen
  randomly from uniform distribution
• Hillclimbing with restarts from best bitstring
  out of K chosen randomly from uniform
  distribution
• Genetic algorithms?

38
COMIT Example of Behavior

• Dummy slide for graphs showing all evaluations in
  TSP domain

39
COMIT Empirical Comparisons

• Big ugly table goes here

40
Summary

• PBIL uses very simple probability distribution
  from which new solutions are generated, yet works
  surprisingly well
• Algorithm using tree-based distributions seems to
  work even better, though at significantly more
  computational expense
• More sophisticated networks past the point of
  diminishing marginal returns? Future research
• COMIT makes tree-based algorithm applicable to
  much larger problems

41
Future Work

• Making algorithm based on complex Bayesian
  Networks more practical
• Combine w/simpler search algorithms, ala COMIT?
• Applying COMIT to more interesting problems
• WALKSAT?
• Optimization in real-valued state spaces. What
  sorts of PDF representations might be useful?
• Gaussians?
• Kernel-based representations?
• Hierarchical representations?

42
Acknowledgements

• Shumeet Baluja
• Doug Baker, Justin Boyan, Lonnie Chrisman, Greg
  Cooper, Geoff Gordon, Andrew Moore, ?