The Spanning Polytope Online Case. Step II

Changing Basis:
Multistage Optimization for
Matroids and Matchings
Anupam Gupta
CMU
Kunal Talwar
Udi Wieder
MSR
MSR
Multi-stage Optimization
β–ͺ Example: A chef needs to buy tomatoes each day, there are 𝑛 suppliers that post
prices every day. Goal: minimize the price paid.
β–ͺ One instance optimization: trivial – buy from the cheapest supplier
β–ͺ Twist: establishing (or breaking) a business relation comes with a fixed cost
β–ͺ Still doable in this case (offline)
β–ͺ More generally, in practice:
β–ͺ As time goes by the optimization problem changes
β–ͺ Change comes with a cost. Either movement cost, setup cost etc.
Multi-stage Optimization
β–ͺ Question: If a single instance could be solved efficiently, can we
solve/approximate well the multi-stage version?
β–ͺ Answer: NO
β–ͺ 3-dimensional matching could be reduced to three instances of a partition matroid
β–ͺ NP-Hard to approximate
β–ͺ Focus of this work: The underlying optimization problem is same matroid
optimization, weight of elements vary in time.
β–ͺ Covers a large class of problem.
β–ͺ Well structured and heavily studied.
Matroid Multi-stage Maintanence (MMM)
Input:
β–ͺ A fixed Matroid (𝐸, 𝑀) of rank π‘Ÿ.
β–ͺ Fixed acquisition cost π‘Ž(𝑒) for acquiring a new item.
β–ͺ Each time step 𝑑, holding costs 𝑐𝑑 (𝑒) for holding an item.
Constraint: β–ͺ Each time step 𝑑, hold a base 𝐡𝑑 βŠ† 𝐸.
Goal:
β–ͺ Minimize total cost:
𝑑𝑐
𝐡𝑑 + π‘Ž 𝐡𝑑 βˆ– π΅π‘‘βˆ’1
β–ͺ Offline version: Both 𝑐 and π‘Ž are given in advance
β–ͺ Online version: π‘Ž is given in advance. 𝑐𝑑 is given at the beginning of time 𝑑.
Matroid Multi-stage Maintanence (MMM)
β–ͺ π‘˜-Uniform Matroid: Base is any set of cardinality π‘˜.
β–ͺ Also known as weighted caching
β–ͺ π‘˜ pages in the cache, each page has a different fetching cost
β–ͺ Graphical Matroids: Independent sets are forests
β–ͺ Each time step hold a spanning tree.
β–ͺ Greedy algorithm solves any single static instance. How to cope with the
acquisition cost?
Related but Different Models
β–ͺ Regret minimization: The cost is given at the end of the time step. In our case, at
the beginning.
β–ͺ Two step optimization: Takes into account two time steps only.
β–ͺ Dynamic Algorithms: Requires optimal solution each time step. Often limits the
amount of change between time steps and aims at minimizing running time.
β–ͺ Example: maintaining a spanning tree with edge inserts and deletions
Related Work
β–ͺ [Buchbinder Chen Naor Shamir 2012]: Same problem, uniform acquisition cost
β–ͺ Provide fractional solutions only.
β–ͺ log 𝑛 approximations
β–ͺ Essentially, an online solution for the Base Polytope LP.
β–ͺ [Bansal Buchbinder Naor 2008] show an integer solution for weighted caching.
β–ͺ Metrical Task Systems – very generic, degenerates to trivial solution in our case –
O(n) approximation.
Main Result
β–ͺ Online algorithm with competitive ratio 𝑂 log 𝐸 log π‘Ÿπ‘‡
β–ͺ Matching lower bound
Time
Universe
size
Matroid
Rank
β–ͺ Hardness of approximation if the underlying problem is maximum matching
The Matroid Polytope and the Base Polytope
β–ͺ The Matroid Polytope is the convex hull of incidence vectors of independent sets
𝐸
β–ͺ Matroid Polytope is π‘₯ ∈ ℝβ‰₯0
| π‘₯ 𝑆 ≀ π‘Ÿπ‘Žπ‘›π‘˜ 𝑆 , βˆ€π‘† βŠ† 𝐸
Sum of π‘₯ in the 𝑆
components
β–ͺ The Base Polytope is the convex hull of all base vectors.
𝐸
β–ͺ Base Polytope = Matroid Polytope ∩ π‘₯ ∈ ℝβ‰₯0
| π‘₯ 𝐸 = π‘Ÿπ‘Žπ‘›π‘˜ 𝐸
First Attempt – Using an LP
β–ͺ Goal: Minimize total cost:
β–ͺ LP: Minimize:
𝑑𝑐
𝑑𝑐
𝐡𝑑 + π‘Ž 𝐡𝑑 βˆ– π΅π‘‘βˆ’1 such that 𝐡𝑑 is a base
𝑋𝑑 + 𝑋𝑑 βˆ’ π‘‹π‘‘βˆ’1
1
such that 𝑋𝑑 in Base Polytope
β–ͺ The LP could be approximated online [Buchbinder Chen Naor Shamir 2012]
β–ͺ But… how to round? Each fractional solution is a convex combination of bases,
would randomized rounding work?
β–ͺ Say we have two vectors 𝑋𝑑 , π‘‹π‘‘βˆ’1 and matching distributions over bases πœ†π‘‘ , πœ†π‘‘βˆ’1 .
β–ͺ The fractional cost is 𝑋𝑑 – π‘‹π‘‘βˆ’1
1,
while the integer cost is 𝐸𝑀(πœ†π‘‘ , πœ†π‘‘βˆ’1 ).
β–ͺ Examples where β„“1 distance is small and EM distance is large suggest that it is hard to
make this approach work.
Key Idea – reduction to spanning sets
β–ͺ A Spanning Set: A set that contains a base = set of full rank
β–ͺ Theorem: Allowing for spanning sets does not reduce the overall cost.
β–ͺ Proof sketch:
β–ͺ We have spanning sets 𝑆1 , 𝑆2 , … , 𝑆𝑇 and want to find bases.
Matroid doesn’t
change!
β–ͺ Set 𝐡1 βŠ† 𝑆1 .
β–ͺ Given π΅π‘‘βˆ’1 and 𝑆𝑑 start with π΅π‘‘βˆ’1 ∩ 𝑆𝑑 and extend it to any base 𝐡𝑑 βŠ† 𝑆𝑑
β–ͺ An item does not drop from 𝐡𝑑 unless it also drops from 𝑆𝑑
β–ͺ Reduction is online
β–ͺ Projection to bases is correlated across time, sidesteps previous issue
Warm-up: Offline Case
β–ͺ Minimize:
𝑑𝑐
𝑋𝑑 + 𝑋𝑑 βˆ’ π‘‹π‘‘βˆ’1
1
such that 𝑋𝑑 in Base Polytope
β–ͺ Pick a random threshold 𝜏 ∈ 0,1 . Round 𝑋𝑑 𝑒 to 1 iff 𝑋𝑑 𝑒 β‰₯ 𝜏 and to 0 o/w
β–ͺ Same expected total cost.
β–ͺ Expected rank of rounded solution is a constant fraction of the rank at each time step.
β–ͺ If repeated 𝑂(log rT) times obtain spanning sets w.h.p. 𝑂(log rT) approximation.
β–ͺ May need to augment the outcome of the rounding but this carries little cost.
β–ͺ There are simpler algorithms (Greedy!) for the offline case.
Online Case. Step I
β–ͺ W.l.o.g each item 𝑒 ∈ 𝐸 has a time interval 𝐼𝑒 such that the item is available at no
cost in 𝐼𝑒 and unavailable outside 𝐼𝑒
β–ͺ Reduction is exact in the offline case and has a constant approximation in the
online case.
β–ͺ Minimize:
Acquisition
Cost
π‘Žπ‘’ π‘₯𝑒
s.t
βˆ€π‘‘, 𝑋𝐸𝑑 ∈ Spanning Polytope
All items available
at time 𝑑
Convex hull of all
spanning sets
The Spanning Polytope
β–ͺ Spanning Polytope of a matroid β„³ is the convex hull of all spanning sets.
β–ͺ Definition: β„³ βˆ— is the dual matroid of β„³. The bases of β„³ βˆ— are the complements
of the bases of β„³.
β–ͺ π‘Ÿβˆ— 𝑆 = π‘Ÿ 𝐸 βˆ– 𝑆 + 𝑆 βˆ’ π‘Ÿ 𝐸
β–ͺ Observation: X in Spanning Polytope of β„³ iff 1 βˆ’ 𝑋 in Matroid Polytope of β„³ βˆ—
Online Case. Step II
β–ͺ Minimize:
π‘Žπ‘’ π‘₯𝑒
Acquisition
Cost
β–ͺ Minimize: π‘Žπ‘’ π‘₯𝑒
s.t
βˆ€π‘‘, π‘₯𝐸𝑑 ∈ Spanning Polytope
All items available
at time 𝑑
Convex hull of all
spanning sets
s.t
β–ͺ βˆ€π‘‘, S βŠ† 𝐸𝑑 π‘₯ 𝑆 β‰₯ π‘Ÿ 𝐸 βˆ’ π‘Ÿ(𝐸 βˆ– 𝑆)
β–ͺ π‘₯𝑒 ∈ [0,1]
β–ͺ This is a Covering LP, that (should be) simpler to solve
Recap
β–ͺ Reduce to β€˜interval instance’.
β–ͺ 𝑂(1) approximation
β–ͺ Solve online: Minimize: π‘Žπ‘’ π‘₯𝑒
β–ͺ 𝑂 log 𝐸 approximation
β–ͺ βˆ€π‘‘, S βŠ† 𝐸𝑑 π‘₯ 𝑆 β‰₯ π‘Ÿ 𝐸 βˆ’ π‘Ÿ(𝐸 βˆ– 𝑆)
β–ͺ π‘₯𝑒 ∈ [0,1]
β–ͺ Round using threshold rounding
β–ͺ 𝑂 log π‘Ÿπ‘‡ approximation
β–ͺ Turn a Spanning Set into a Base
β–ͺ Exact
β–ͺ 𝑂 log 𝐸 log π‘Ÿπ‘‡ approximation
β–ͺ Hardness: Ξ© log 𝐸 log 𝑇
Solving the LP
β–ͺ Minimize: π‘Žπ‘’ π‘₯𝑒
β–ͺ βˆ€π‘‘, S βŠ† 𝐸𝑑
π‘₯ 𝑆 β‰₯ π‘Ÿ 𝐸 βˆ’ π‘Ÿ(𝐸 βˆ– 𝑆)
β–ͺ π‘₯𝑒 ∈ [0,1]
β–ͺ [Buchbinder Naor] provide a method for solving these types of LP’s online
β–ͺ Can we use it in a Black Box way?
β–ͺ Yes but… we have an exponential number of constraints
β–ͺ Overcome the problem by invoking [BN] on adaptively selected sets of violated
constraints.
β–ͺ Show that running time stays polynomial
Hardness of Perfect Matching
β–ͺ We are given a graph 𝐺 = (𝑉, 𝐸) and need to maintain a perfect matching each
time step.
β–ͺ Theorem: NP-Hard to distinguish between instances with cost 𝑉
β–ͺ Even when holding costs are in 0, ∞
β–ͺ Acquisition costs are uniform for all edges.
β–ͺ Number of time steps is a small constant.
πœ–
and 𝑉
1βˆ’πœ– .
Concluding Remarks
β–ͺ Better algorithms for the uniform cost case.
β–ͺ Greedy?
β–ͺ Hardness/upper-bound for intersections of matroids.
β–ͺ Matchings in bipartite graphs.