Content deleted Content added
link LIBSVM and improve reference |
No edit summary Tags: Mobile edit Mobile web edit |
||
| (30 intermediate revisions by 19 users not shown) | |||
Line 1:
{{short description|Algorithm for solving the quadratic programming problem from training SVMs}}
'''Sequential minimal optimization''' ('''SMO''') is an algorithm for efficiently solving the optimization problem which arises during the training of [[support vector machine]]s. It was invented by [[John Platt (Principal Researcher)|John Platt]] in 1998 at [[Microsoft Research]].<ref>{{Citation▼
{{Infobox Algorithm
|image=
|class=[[Optimization algorithm]] for training support vector machines
|data=
|time=O(''n''³)
|space=
}}
▲'''Sequential minimal optimization''' ('''SMO''') is an algorithm for
| last = Platt | first = John
| year = 1998
| title = Sequential Minimal Optimization: A Fast Algorithm for Training Support Vector Machines
|
| url = https://www.microsoft.com/en-us/research/wp-content/uploads/2016/02/tr-98-14.pdf
}}</ref> SMO is widely used for training support vector machines and is implemented by the popular [[LIBSVM]] tool.<ref>{{cite journal
|last1=Chang |first1=Chih-Chung
Line 10 ⟶ 19:
|journal=ACM Transactions on Intelligent Systems and Technology
|volume=2 |issue=3 |year=2011
|doi=10.1145/1961189.1961199
}}</ref><ref>Luca Zanni (2006). ''[http://jmlr.csail.mit.edu/papers/volume7/zanni06a/zanni06a.pdf Parallel Software for Training Large Scale Support Vector Machines on Multiprocessor Systems]''.</ref> The publication of the SMO algorithm in 1998 has generated a lot of excitement in the SVM community, as previously available methods for SVM training were much more complex and required expensive third-party [[Quadratic programming|QP]] solvers.<ref>{{Citation▼
|s2cid=961425
▲}}</ref><ref>{{cite web |first=Luca |last=Zanni
| last = Rifkin | first = Ryan
| year = 2002
|
| title = Everything Old is New Again: a Fresh Look at Historical Approaches in Machine Learning
| type = Ph.D. Thesis |publisher=Massachusetts Institute of Technology
| pages = 18
}}</ref>
Line 34 ⟶ 45:
:<math>0 \leq \alpha_1, \alpha_2 \leq C,</math>
:<math>y_1 \alpha_1 + y_2 \alpha_2 = k,</math>
and this reduced problem can be solved analytically: one needs to find a minimum of a one-dimensional quadratic function. <math>k</math> is the negative of the sum over the rest of terms in the equality constraint, which is fixed in each iteration.
The algorithm proceeds as follows:
Line 44 ⟶ 55:
# Repeat steps 1 and 2 until convergence.
When all the Lagrange multipliers satisfy the KKT conditions (within a user-defined tolerance), the problem has been solved. Although this algorithm is guaranteed to converge, heuristics are used to choose the pair of multipliers so as to accelerate the rate of convergence. This is critical for large data sets since there are <math>n(n-1)/2</math> possible choices for <math>\alpha_i</math> and <math>\alpha_j</math>.
==
The first approach to splitting large SVM learning problems into a series of smaller optimization tasks was proposed by [[Bernhard Boser]], [[Isabelle Guyon]], and [[Vladimir Vapnik]].<ref name="ReferenceA">{{Cite book | doi = 10.1145/130385.130401| chapter = A training algorithm for optimal margin classifiers| title = Proceedings of the fifth annual workshop on Computational learning theory - COLT '92| pages = 144| year = 1992| last1 = Boser | first1 = B. E. | last2 = Guyon | first2 = I. M. | last3 = Vapnik | first3 = V. N. | isbn = 978-0897914970| citeseerx = 10.1.1.21.3818| s2cid = 207165665}}</ref> It is known as the "chunking algorithm". The algorithm starts with a random subset of the data, solves this problem, and iteratively adds examples which violate the optimality conditions. One disadvantage of this algorithm is that it is necessary to solve QP-problems scaling with the number of SVs. On real world sparse data sets, SMO can be more than 1000 times faster than the chunking algorithm.<ref name = "Platt"/>
{{reflist}}▼
In 1997, [[E. Osuna]], [[R. Freund]], and [[F. Girosi]] proved a theorem which suggests a whole new set of QP algorithms for SVMs.<ref>{{Cite book | doi = 10.1109/NNSP.1997.622408| chapter = An improved training algorithm for support vector machines| title = Neural Networks for Signal Processing [1997] VII. Proceedings of the 1997 IEEE Workshop| pages = 276–285| year = 1997| last1 = Osuna | first1 = E. | last2 = Freund | first2 = R. | last3 = Girosi | first3 = F. | isbn = 978-0-7803-4256-9| citeseerx = 10.1.1.392.7405| s2cid = 5667586}}</ref> By the virtue of this theorem a large QP problem can be broken down into a series of smaller QP sub-problems. A sequence of QP sub-problems that always add at least one violator of the [[Karush–Kuhn–Tucker conditions|Karush–Kuhn–Tucker (KKT) conditions]] is guaranteed to converge. The chunking algorithm obeys the conditions of the theorem, and hence will converge.<ref name = "Platt"/> The SMO algorithm can be considered a special case of the Osuna algorithm, where the size of the optimization is two and both Lagrange multipliers are replaced at every step with new multipliers that are chosen via good heuristics.<ref name = "Platt"/>
The SMO algorithm is closely related to a family of optimization algorithms called [[Bregman method]]s or row-action methods. These methods solve convex programming problems with linear constraints. They are iterative methods where each step projects the current primal point onto each constraint.<ref name = "Platt"/>
== See also ==
* [[Kernel perceptron]]
== References ==
▲{{reflist|30em}}
{{Optimization algorithms}}
[[Category:Optimization algorithms and methods]]
[[Category:Support vector machines]]
| |||