Dynamic Meta-Learning
Matteo Gagliolo
IDSIA, Galleria 2, 6928 Manno (Lugano), Switzerland
University of Lugano, Faculty of Informatics,
Via Buffi 13, 6904 Lugano, Switzerland
Most solvable AI problems can be addressed by more than one algorithm; most
AI algorithms feature a number of parameters that have to be set. Both choices can
dramatically affect the quality of the obtained solution, and the time spent obtaining
it. Algorithm Selection, or Meta-Learning, techniques [1, 2] typically address these
questions by solving a large number of problems with each of the available algorithms,
in order to learn a mapping from (problem,algorithm) pairs to expected performance.
The obtained mapping is later used to select and run, for each new problem, only the
algorithm that is expected to give the best results.
This approach, tough being preferable to the far more popular “trial and error”,
poses a number of problems. It presumes that such a mapping can be learned at all,
i.e., that the actual algorithm performance on a given problem will be predictable with
enough precision before even starting the algorithm — often not the case with stochas-
tic algorithms, whose performance can exhibit large fluctuations among different runs
(see, e.g., [3] ). It also assumes problem instances met during the training phase to
be statistically representative of successive ones. For these reasons, there usually is
no way to detect a relevant discrepancy between expected and actual performance of
the chosen algorithm. Finally, it neglects computational complexity issues: ranking
between algorithms is often based solely on the expected quality of the performance;
and the time spent during the training phase is not even considered, although it can
be large enough to cancel any practical advantage of algorithm selection.
The Algorithm Portfolio paradigm [4, 5] consists in selecting a subset of the avail-
able algorithms, to be run in parallel, with the same priority, until the fastest one
solves the problem. This simple scheme is more robust, as it’s less likely that per-
formance estimates will be wrong for all selected algorithms, but it also involves
an additional overhead, due to the “brute force” parallel execution of all candidate
solvers.
In our view, a crucial weakness of these approaches is that they don’t exploit any
feedback from the actual execution of the chosen algorithms. We tried to move a
step in this direction, introducing Dynamic Algorithm Portfolios [6, 7]. Instead of
first choosing a portfolio then running it, we iteratively allocate a time slice, shar-
ing it among all the available algorithms, and update the relative priorities of the
algorithms, based on their current state, in order to favor the most promising ones.
Instead of basing the priority attribution on performance quality, we fix a target per-
formance, and minimize the time to reach it. To this aim, we search for a mapping
from (problem,algorithm,current algorithm state) triples to expected time to reach the
desired performance quality. The mapping is obtained training a parametric model
of algorithm runtime distribution. To further reduce computational complexity, we
1

focus on lifelong-learning techniques that drop the artificial boundary between train-
ing and usage, exploiting the mapping during training, and including training time in
performance evaluation.
The obtained selection technique is generic, not depending on algorithm-specific
properties. We present experiments with genetic algorithms [8] and satisfiability prob-
lem solvers [9].
The target of our work is to obtain a fully dynamic meta-learning agent that learns
to use a set of algorithms by solving a set of problems, with minimal a-priori knowl-
edge, and minimal performance overhead, allowed by a continuous cycle of runtime
feedback and re-allocation of computational resources.
References
[1] J. R. Rice, “The algorithm selection problem,” in Advances in computers (M. Ru-
binoff and M. C. Yovits, eds.), vol. 15, pp. 65–118, New York: Academic Press,
1976.
[2] R. Vilalta and Y. Drissi, “A perspective view and survey of meta-learning,” Artif.
Intell. Rev., vol. 18, no. 2, pp. 77–95, 2002.
[3] C. P. Gomes, B. Selman, N. Crato, and H. Kautz, “Heavy-tailed phenomena in
satisfiability and constraint satisfaction problems,” J. Autom. Reason., vol. 24,
no. 1-2, pp. 67–100, 2000.
[4] B. A. Huberman, R. M. Lukose, and T. Hogg, “An economic approach to hard
computational problems,” Science, vol. 275, pp. 51–54, 1997.
[5] C. P. Gomes and B. Selman, “Algorithm portfolios,” Artificial Intelligence,
vol. 126, no. 1–2, pp. 43–62, 2001.
[6] M. Gagliolo and J. Schmidhuber, “Dynamic algorithm portfolios,” in Ninth Inter-
national Symposium on Artificial Intelligence and Mathematics, Fort Lauderdale,
Florida, January 2006, 2006.
[7] M. Gagliolo and J. Schmidhuber, “A neural network model for inter-problem adap-
tive online time allocation,” in ICANN 2005 Proceedings, Part 2 (W. Duch et al.,
eds.), pp. 7–12, Springer, 2005.
[8] J. H. Holland, Adaptation in Natural and Artificial Systems. University of Michi-
gan Press, Ann Arbor, 1975.
[9] J. Gu, P. W. Purdom, J. Franco, and B. W. Wah, “Algorithms for the satisfiability
(SAT) Problem: A survey,” in Satisfiability Problem: Theory and applications (D.-
Z. Du et al., eds.), pp. 19–152, American Mathematical Society, 1997.