Hybrid Evolution of Heterogeneous Neural Networks
Transcription
Hybrid Evolution of Heterogeneous Neural Networks
Hybrid Evolution of Heterogeneous Neural Networks 01001110 01100101 01110101 01110010 01101111 01101110 01101111 01110110 01100001 00100000 01110011 01101011 01110101 01110000 01101001 01101110 01100001 00100000 01101011 01100001 01110100 01100101 01100100 01110010 01111001 00100000 01110000 01101111 01100011 01101001 01110100 01100001 01100011 01110101 00101100 00100000 01000110 01000101 01001100 00100000 01000011 01010110 01010101 01010100 Zdeněk Buk [email protected] Miroslav Šnorek [email protected] http://cig.felk.cvut.cz Computational Intelligence Group Department of Computer Science and Engineering Faculty of Electrical Engineering Czech Technical University in Prague ICANN 2008 Outline ● Continual Evolution Algorithm (CEA) description ● Data structures – encoding of the individuals ● Evolution process ● Control functions ● Testing, experiments ● Population behavior ● Conclusion 18th International Conference on Artificial Neural Networks 3-6 September 2008, Diplomat Hotel, Prague, Czech Republic Zdeněk Buk, [email protected], http://cig.felk.cvut.cz Continual Evolution Algorithm ● Hybrid genetic algorithm – ● combination of genetic and gradient-based methods Separate evolution of structure and parameters of individuals (models, neural networks) ● Variable population size ● Sequential replacement of individuals – evolution in continual time dimension – age parameter of each individual 18th International Conference on Artificial Neural Networks 3-6 September 2008, Diplomat Hotel, Prague, Czech Republic Zdeněk Buk, [email protected], http://cig.felk.cvut.cz Individuals encoding xi =a i , pi , si , bi ● Separate evolution of structure and parameters of individuals (models, neural networks) 18th International Conference on Artificial Neural Networks 3-6 September 2008, Diplomat Hotel, Prague, Czech Republic Zdeněk Buk, [email protected], http://cig.felk.cvut.cz Individuals encoding xi =a i , pi , si , bi age of the individual 18th International Conference on Artificial Neural Networks 3-6 September 2008, Diplomat Hotel, Prague, Czech Republic Zdeněk Buk, [email protected], http://cig.felk.cvut.cz Individuals encoding xi =a i , pi , si , bi structural vector 18th International Conference on Artificial Neural Networks 3-6 September 2008, Diplomat Hotel, Prague, Czech Republic Zdeněk Buk, [email protected], http://cig.felk.cvut.cz Individuals encoding xi =a i , pi , si , bi structural vector topology of the network 18th International Conference on Artificial Neural Networks 3-6 September 2008, Diplomat Hotel, Prague, Czech Republic Zdeněk Buk, [email protected], http://cig.felk.cvut.cz Individuals encoding xi =a i , pi , si , bi structural vector topology of the network activation functions 18th International Conference on Artificial Neural Networks 3-6 September 2008, Diplomat Hotel, Prague, Czech Republic Zdeněk Buk, [email protected], http://cig.felk.cvut.cz Individuals encoding xi =a i , pi , si , bi parametric vector behavioral vector 18th International Conference on Artificial Neural Networks 3-6 September 2008, Diplomat Hotel, Prague, Czech Republic Zdeněk Buk, [email protected], http://cig.felk.cvut.cz Individuals encoding xi =a i , pi , si , bi parametric vector behavioral vector weights 18th International Conference on Artificial Neural Networks 3-6 September 2008, Diplomat Hotel, Prague, Czech Republic Zdeněk Buk, [email protected], http://cig.felk.cvut.cz Individuals encoding xi =a i , pi , si , bi 18th International Conference on Artificial Neural Networks 3-6 September 2008, Diplomat Hotel, Prague, Czech Republic Zdeněk Buk, [email protected], http://cig.felk.cvut.cz Evolution process ● Hybrid genetic algorithm – ● combination of genetic and gradient-based methods 2 Dimensional evolution 18th International Conference on Artificial Neural Networks 3-6 September 2008, Diplomat Hotel, Prague, Czech Republic Zdeněk Buk, [email protected], http://cig.felk.cvut.cz Evolution process 18th International Conference on Artificial Neural Networks 3-6 September 2008, Diplomat Hotel, Prague, Czech Republic Zdeněk Buk, [email protected], http://cig.felk.cvut.cz Crossover and mutation Individual 1 p1 s1 b1 p2 s2 b2 Offspring Individual 2 18th International Conference on Artificial Neural Networks 3-6 September 2008, Diplomat Hotel, Prague, Czech Republic Zdeněk Buk, [email protected], http://cig.felk.cvut.cz Crossover and mutation Individual 1 p1 s1 b1 p2 s2 b2 Random mutation Offspring Individual 2 18th International Conference on Artificial Neural Networks 3-6 September 2008, Diplomat Hotel, Prague, Czech Republic Zdeněk Buk, [email protected], http://cig.felk.cvut.cz Crossover and mutation Individual 1 p1 s1 b1 p2 s2 b2 Random mutation Offspring Individual 2 18th International Conference on Artificial Neural Networks 3-6 September 2008, Diplomat Hotel, Prague, Czech Republic Zdeněk Buk, [email protected], http://cig.felk.cvut.cz Crossover and mutation Individual 1 p1 s1 b1 p2 s2 b2 Offspring Individual 2 18th International Conference on Artificial Neural Networks 3-6 September 2008, Diplomat Hotel, Prague, Czech Republic Zdeněk Buk, [email protected], http://cig.felk.cvut.cz Crossover and mutation Individual 1 p1 s1 b1 p2 s2 b2 Offspring Individual 2 18th International Conference on Artificial Neural Networks 3-6 September 2008, Diplomat Hotel, Prague, Czech Republic Zdeněk Buk, [email protected], http://cig.felk.cvut.cz Crossover and mutation Individual 1 p1 s1 b1 p2 s2 b2 Offspring Individual 2 Random mutation 18th International Conference on Artificial Neural Networks 3-6 September 2008, Diplomat Hotel, Prague, Czech Republic Zdeněk Buk, [email protected], http://cig.felk.cvut.cz Crossover and mutation Individual 1 p1 s1 b1 p2 s2 b2 Offspring Copy Individual 2 18th International Conference on Artificial Neural Networks 3-6 September 2008, Diplomat Hotel, Prague, Czech Republic Zdeněk Buk, [email protected], http://cig.felk.cvut.cz Crossover and mutation Individual 1 p1 s1 b1 p2 s2 b2 Offspring pi si bi Individual 2 18th International Conference on Artificial Neural Networks 3-6 September 2008, Diplomat Hotel, Prague, Czech Republic Zdeněk Buk, [email protected], http://cig.felk.cvut.cz Time dimension ● Training the behavioral vector during time using gradient algorithm. Age=0 pi si bi Time 18th International Conference on Artificial Neural Networks 3-6 September 2008, Diplomat Hotel, Prague, Czech Republic Zdeněk Buk, [email protected], http://cig.felk.cvut.cz Time dimension ● Training the behavioral vector during time using gradient algorithm. Age=0 Age=1 pi si bi Time 18th International Conference on Artificial Neural Networks 3-6 September 2008, Diplomat Hotel, Prague, Czech Republic Zdeněk Buk, [email protected], http://cig.felk.cvut.cz Time dimension ● Training the behavioral vector during time using gradient algorithm. Age=0 Age=1 Age=2 pi si bi Time 18th International Conference on Artificial Neural Networks 3-6 September 2008, Diplomat Hotel, Prague, Czech Republic Zdeněk Buk, [email protected], http://cig.felk.cvut.cz Probability control ● Probability functions ● RP – reproduction probability, ● DP – death probability, elimination of bad solutions in population. ● Balancing functions ● to keep the population size in some reasonable limits. 18th International Conference on Artificial Neural Networks 3-6 September 2008, Diplomat Hotel, Prague, Czech Republic Zdeněk Buk, [email protected], http://cig.felk.cvut.cz Reproduction probability ● Raw reproduction probability function * * RP xi = RP a i , F xi Fitness RP* Age Fitness Age 18th International Conference on Artificial Neural Networks 3-6 September 2008, Diplomat Hotel, Prague, Czech Republic Zdeněk Buk, [email protected], http://cig.felk.cvut.cz Death probability ● Raw death probability function * * DP xi = DP a i , F xi Fitness DP* Age Fitness Age 18th International Conference on Artificial Neural Networks 3-6 September 2008, Diplomat Hotel, Prague, Czech Republic Zdeněk Buk, [email protected], http://cig.felk.cvut.cz Balancing functions ● Computation of final probabilities. ● Depend on raw probabilities and population size. ● Big population grows slower, ● small population grows faster. * RP xi = BAL RP N , RP xi * DP xi = BAL DP N , DP xi 18th International Conference on Artificial Neural Networks 3-6 September 2008, Diplomat Hotel, Prague, Czech Republic Zdeněk Buk, [email protected], http://cig.felk.cvut.cz Evolution control 18th International Conference on Artificial Neural Networks 3-6 September 2008, Diplomat Hotel, Prague, Czech Republic Zdeněk Buk, [email protected], http://cig.felk.cvut.cz Evolution control 18th International Conference on Artificial Neural Networks 3-6 September 2008, Diplomat Hotel, Prague, Czech Republic Zdeněk Buk, [email protected], http://cig.felk.cvut.cz Evolution control 18th International Conference on Artificial Neural Networks 3-6 September 2008, Diplomat Hotel, Prague, Czech Republic Zdeněk Buk, [email protected], http://cig.felk.cvut.cz Evolution control 18th International Conference on Artificial Neural Networks 3-6 September 2008, Diplomat Hotel, Prague, Czech Republic Zdeněk Buk, [email protected], http://cig.felk.cvut.cz Evolution control 18th International Conference on Artificial Neural Networks 3-6 September 2008, Diplomat Hotel, Prague, Czech Republic Zdeněk Buk, [email protected], http://cig.felk.cvut.cz Evolution control 18th International Conference on Artificial Neural Networks 3-6 September 2008, Diplomat Hotel, Prague, Czech Republic Zdeněk Buk, [email protected], http://cig.felk.cvut.cz Evolution control 18th International Conference on Artificial Neural Networks 3-6 September 2008, Diplomat Hotel, Prague, Czech Republic Zdeněk Buk, [email protected], http://cig.felk.cvut.cz Testing, experiments ● ● ● ● Construction of the neural networks. Universal topology based on fully recurrent network. Structure – topology → adjacency matrix – activation functions → parametrized Λ-functions Behavior – weight matrix 18th International Conference on Artificial Neural Networks 3-6 September 2008, Diplomat Hotel, Prague, Czech Republic Zdeněk Buk, [email protected], http://cig.felk.cvut.cz Testing, experiments ● ● Neural networks based on fully recurrent topology construction using CEA Activation function optimization 18th International Conference on Artificial Neural Networks 3-6 September 2008, Diplomat Hotel, Prague, Czech Republic Zdeněk Buk, [email protected], http://cig.felk.cvut.cz Testing, experiments ● Benchmark tasks ● Learn to oscillate experiment 18th International Conference on Artificial Neural Networks 3-6 September 2008, Diplomat Hotel, Prague, Czech Republic Zdeněk Buk, [email protected], http://cig.felk.cvut.cz Population behavior 18th International Conference on Artificial Neural Networks 3-6 September 2008, Diplomat Hotel, Prague, Czech Republic Zdeněk Buk, [email protected], http://cig.felk.cvut.cz Population behavior 18th International Conference on Artificial Neural Networks 3-6 September 2008, Diplomat Hotel, Prague, Czech Republic Zdeněk Buk, [email protected], http://cig.felk.cvut.cz Fitness Population behavior Iterations 18th International Conference on Artificial Neural Networks 3-6 September 2008, Diplomat Hotel, Prague, Czech Republic Zdeněk Buk, [email protected], http://cig.felk.cvut.cz Conclusion ● ● Reduction of the number of fitness function evaluations. Using the probability and balancing functions it is possible to change the CEA to behave more – – – like random search, standard genetic algorithm, or gradient algorithm. 18th International Conference on Artificial Neural Networks 3-6 September 2008, Diplomat Hotel, Prague, Czech Republic Zdeněk Buk, [email protected], http://cig.felk.cvut.cz Conclusion ● CEA → universal optimization algorithm ● mainly for problems with separate description – of structure – and behavior → typically neural networks. ● Automatic control of the size of the population – exploitation ↔ exploration 18th International Conference on Artificial Neural Networks 3-6 September 2008, Diplomat Hotel, Prague, Czech Republic Zdeněk Buk, [email protected], http://cig.felk.cvut.cz