论文信息 - Towards Life-Long Meta Learning

Towards Life-Long Meta Learning

We reformulate algorithm selection as a time allocationproblem: all candidate algorithms are run in parallel, and their relativ e priorities are continually updated based on its current time to solution, e stimated according to a parametric model that is trained and used while solving a sequence of problems.

Matteo Gagliolo | Juergen Schmidhuber

[1] David Maxwell Chickering,et al. A Bayesian Approach to Tackling Hard Computational Problems (Preliminary Report) , 2001, Electron. Notes Discret. Math..

[2] Shlomo Zilberstein,et al. Monitoring and control of anytime algorithms: A dynamic programming approach , 2001, Artif. Intell..

[3] Marek Petrik,et al. Statistically Optimal Combination of Algorithms , 2004 .

[4] Jürgen Schmidhuber,et al. Adaptive Online Time Allocation to Search Algorithms , 2004, ECML.

[5] Jürgen Schmidhuber,et al. A Neural Network Model for Inter-problem Adaptive Online Time Allocation , 2005, ICANN.

[6] Michail G. Lagoudakis,et al. Reinforcement Learning for Algorithm Selection , 2000, AAAI/IAAI.

[7] Michail G. Lagoudakis,et al. Algorithm Selection using Reinforcement Learning , 2000, ICML.

[8] Bart Selman,et al. Algorithm portfolios , 2001, Artif. Intell..

[9] F. Lobo,et al. A parameter-less genetic , 1999 .