BRAINSTORMING

Agent based Meta-learning Approach

Dariusz Plewczynski

ICM, Interdisciplinary Centre for Mathematical and Computational Modelling, University of Warsaw

Pawinskiego 5a Street, 02-106 Warsaw, Poland

Keywords: Machine learning, Ensemble methods, Agent based modelling, Meta-learning, Brainstorming.

Abstract: Brainstorming meta-learning approach is performed by a set of cognitive agents (CA), each implementing

different machine learning (ML) algorithm, and/or trained using diverse subsets of available features

describing input examples. The goal of the meta-learning procedure is providing a general and flexible

classification meta-model for a given training data. In the first phase all agents, when trained using different

features describing training objects, construct the ensemble of classification models independently. In the

second step all solutions are gathered and the consensus is built between them by optimizing the voting

weights for all agents. No early solution, given even by a generally low performing agent, is not discarded

until the late phase of prediction, when comparing different learning models draws the final conclusion. The

final phase, i.e. brainstorming tries to balance the generality of solution and the overall cognitive

performance of all CAs. The classification meta-model is than used for predictions of the classification

membership for given testing examples. The method was recently used in several ML applications in

bioinformatics and chemoinformatics by the author.

1 INTRODUCTION

Although quite a few machine learning algorithms

have been developed, such as fuzzy or crisp

clustering, support vector machine (SVM), artificial

neural network (ANN), or K-nearest neighbor

(KNN), and many others classifiers, the way they

operate the identification is basically individual. Yet,

the proper approach usually takes into account the

opinions from several experts rather than rely on

only one when they are making critical decisions.

Likewise, a sophisticated identifier should be trained

by several different modes. This is the core idea of

brainstorming, i.e. the consensus meta-learning

algorithm that will be described in this manuscript,

which applications were described previously for

selected types of machine learning methods such as

clustering (Can, T. et al., 2004), (Han, X., 2007),

(Schulze-Kremer and King, 1992), (Vernikos and

Parkhill, 2008), support vector machine (Arimoto et

al., 2005), (Bhasin and Raghava, 2004), (Briem and

Gunther, 2005), (Burton, J. et al., 2009), (Yao, X. Q.

et al., 2008), (Abrusan, G. et al., 2009), (Hwang, S.

et al., 2007), artificial neural networks (Yao and

She, 2008), (Garg, P. et al., 2009), (Miller and

Blom., 2009), or K-nearest neighbor (Arimoto et al.,

2005), (Briem and Gunther, 2005), (Garg, P. et al.,

2009), (Bindewald and Shapiro, 2006), including our

own findings (Plewczynski and Ginalski, 2009),

(Plewczynski, D. et al., 2007), (Plewczynski, D. et

al., 2009a), (Plewczynski, D. et al., 2005),

(Plewczynski, D. et al. 2009b), (Plewczynski, D. et

al., 2006), Bagging and boosting are previously

known meta-learning techniques had a wide array of

applications as recapitulated in various manuscripts

(Bruce, C. L. et al., 2007), (Islam, M. M. et al.,

2008), (Plewczynski, D. et al., 2009c), (Schwenk

and Bengio, 2000), (Serpen, G. et al., 2008),

(Shrestha and Solomatine, 2006), (Peng, Y., 2006),

(Wang, C. W., 2006), (Yang, J. Y. et al., 2008), The

meta-learning procedure is here implemented using

Agent Based Modelling framework developed

recently for various applications in Life Sciences

(Devillers, J. et al.), (Moore, D. et al., 2009),

(Farmer and Foley,2009), (Gu and Novak, 2009),

(Pogson, M. et al., 2008), (Sun, T. et al., 2008),

(Bryson, J. J. et al., 2007), (Pogson, M. et al., 2006),

(Walker, D. C. et al., 2004).

487

Plewczynski D..

BRAINSTORMING - Agent based Meta-learning Approach.

DOI: 10.5220/0003310104870490

In Proceedings of the 3rd International Conference on Agents and Artiﬁcial Intelligence (ICAART-2011), pages 487-490

ISBN: 978-989-8425-41-6

 2011 SCITEPRESS (Science and Technology Publications, Lda.)

2 CONSENSUS OF MACHINE

LEARNING METHODS

The model of meta-learning within Agent Based

Modelling (ABM) framework is based on several

assumptions:

1. Binary Logic

I assume the binary logic of individual

intelligent, or so called cognitive agents CAs, i.e. we

deal with N different software agents. For the single

prediction, each algorithm gives one of two opposite

decisions (“YES” or “NO”) described here by the

variable

=±

. Typically ML algorithms, such as

support vector machines, decision trees, trend

vectors, artificial neural networks, random forest,

predict two classes for incoming data, based on

trained models. Therefore the prediction of all ML

algorithms addresses the same question: is a query

item in the class of positives (“YES”), or it is not

(“NO”).

2. Strength Parameters

Each CA is characterized typically by two

parameters:

(

)

,p f precision j

and

(

)

f recall j

that

describe the quality of predictions for individual

algorithm implemented by an agent (described by j

index). It depends on a training dataset; the values of

those parameters will be different for each of

training session, or cognitive task. Therefore the

parameters should be averaged over different

cognitive tasks in order to make them data-

independent. The quality of brainstorming approach

depends on mean values

∑

and

∑

calculated over used learners.

3. Probability of Success

The weighted majority-minority balance in the

system is given by the equation:

(

)

()

jjj

Ns p

∑

(1)

The normalized and nonnegative value of m

describes the probability for the correct prediction,

i.e. we assume here the modified or weighted vote

rule. Each learner votes for the final prediction

outcome, all votes are gathered, and the relative

probability of correct answer is calculated, as given

by the set of individual learners.

4. Brainstorming: The Procedure of Meta-learning

The global preference toward selected solution in

brainstorming method is described as the global

order parameter that is calculated using all used

CAs. Each cognitive network node (learner, or in

other words intelligent agent) performs its own and

independent training on available input data (both

the training and testing datasets are identical for all

learners). In the prediction step, a query of testing

items is analyzed independently by each agent,

which predicts the query item classification (positive

or negative). Then, all predictions done by a set of

learners are gathered and integrated into the single

prediction via majority rule. This view of the

consensus as between various machine learning

algorithms is especially useful for artificial

intelligence, or robotic applications, where adaptive

behavior given by the integration of results from a

set of ML methods.

3 CONCLUSIONS

Generally there are two competing philosophies in

supervised learning, where goal is to minimize the

probability of model errors on future data. A single

model approach tries to build a single good model:

either not using Occam’s razor principle (Minimax

Probability Machine, trees, Neural Networks,

Nearest Neighbor, Radial Basis Functions) or those

based on Occam’s razor models that select the best

model as the simplest one (Support Vector

Machines, Bayesian Methods, other kernel based

methods such as Kernel Matching Pursuit). An

ensemble of models states that a good single model

is difficult to compute, so it tries to build many of

those and combine them. Combining many

uncorrelated models produces better predictors as

was observed in models that don’t use randomness

or use directed randomness (Boosting, Specific cost

function, Gradient Boosting, a boosting algorithm

derivative for any cost function), or in models that

incorporate randomness (Bagging, Bootstrap

Sample: Uniform random sampling with

replacement, Stochastic Gradient Boosting, Random

Forests, or by inputs randomizations for splitting at

tree nodes).

Meta-learning approach trains an ensemble of

machine learning algorithms on the whole or

different subset of all available training examples.

The consensus gathers all solutions and tries to

balance between them in order to maximize the

prediction performance. No early solution, even

provided by a generally low performing module, is

not discarded until the late phase of prediction, when

ICAART 2011 - 3rd International Conference on Agents and Artificial Intelligence

488

comparing different machine learning classifiers

draws the final conclusion. This final phase is

focusing on balance the generality of solution and

the overall performance of trained model. Early

results shows, that brainstorming approach reaches

higher performance than any single method used in

consensus. This confirms reported results of other

meta-learning approaches based on different

versions of single machine learning algorithm or

those that use a set of different ML (Plewczynski,

D., 2009), (Plewczynski, D., 1998), (Plewczynski,

D., 2010).

ACKNOWLEDGEMENTS

This work was supported by Polish Ministry of

Education and Science (N301 159735, N518 409238

and others).

REFERENCES

Can, T., et al., Automated protein classification using

consensus decision. Proc IEEE Comput Syst

Bioinform Conf, 2004: p. 224-35.

Han, X., Cancer molecular pattern discovery by subspace

consensus kernel classification. Comput Syst

Bioinformatics Conf, 2007. 6: p. 55-65.

Schulze-Kremer, S. and R.D. King, IPSA-Inductive

Protein Structure Analysis. Protein Eng, 1992. 5(5): p.

377-90.

Vernikos, G. S. and J. Parkhill, Resolving the structural

features of genomic islands: a machine learning

approach. Genome Res, 2008. 18(2): p. 331-42.

Arimoto, R., M. A. Prasad, and E.M. Gifford,

Development of CYP3A4 inhibition models:

comparisons of machine-learning techniques and

molecular descriptors. J Biomol Screen, 2005. 10(3):

p. 197-205.

Bhasin, M. and G. P. Raghava, Prediction of CTL epitopes

using QM, SVM and ANN techniques. Vaccine, 2004.

22(23-24): p. 3195-204.

Briem, H. and J. Gunther, Classifying "kinase inhibitor-

likeness" by using machine-learning methods.

Chembiochem, 2005. 6(3): p. 558-66.

Burton, J., et al., Virtual screening for cytochromes p450:

successes of machine learning filters. Comb Chem

High Throughput Screen, 2009. 12(4): p. 369-82.

Yao, X. Q., H. Zhu, and Z. S. She, A dynamic Bayesian

network approach to protein secondary structure

prediction. BMC Bioinformatics, 2008. 9: p. 49.

Abrusan, G., et al., TEclass--a tool for automated

classification of unknown eukaryotic transposable

elements. Bioinformatics, 2009. 25(10): p. 1329-30.

Hwang, S., Z. Gou, and I. B. Kuznetsov, DP-Bind: a web

server for sequence-based prediction of DNA-binding

residues in DNA-binding proteins. Bioinformatics,

2007. 23(5): p. 634-6.

Garg, P., et al., SubCellProt: predicting protein

subcellular localization using machine learning

approaches. In Silico Biol, 2009. 9(1-2): p. 35-44.

Miller, M. L. and N. Blom, Kinase-specific prediction of

protein phosphorylation sites. Methods Mol Biol,

2009. 527: p. 299-310, x.

Bindewald, E. and B. A. Shapiro, RNA secondary

structure prediction from sequence alignments using a

network of k-nearest neighbor classifiers. RNA, 2006.

12(3): p. 342-52.

Plewczynski, D. and K. Ginalski, The interactome:

predicting the protein-protein interactions in cells.

Cell Mol Biol Lett, 2009. 14(1): p. 1-22.

Plewczynski, D., et al., The RPSP: Web server for

prediction of signal peptides. Polymer, 2007. 48: p.

5493-5496.

Plewczynski, D., A. H. Spieser, and U. Koch,

Performance of Machine Learning Methods for

Ligand-Based Virtual Screening. Combinatorial

Chemistry & High Throughput Screening, 2009.

Plewczynski, D., et al., AutoMotif server: prediction of

single residue post-translational modifications in

proteins. Bioinformatics, 2005. 21(10): p. 2525-7.

Plewczynski, D., et al., Virtual High Throughput

Screening Using Combined Random Forest and

Flexible Docking. Combinatorial Chemistry & High

Throughput Screening, 2009.

Plewczynski, D., S. A. H. Spieser, and U. Koch, Assessing

different classification methods for virtual screening.

Journal of Chemical Information and Modeling, 2006.

46(3): p. 1098-1106.

Bruce, C. L., et al., Contemporary QSAR classifiers

compared. J Chem Inf Model, 2007. 47(1): p. 219-27.

Islam, M. M., et al., Bagging and boosting negatively

correlated neural networks. IEEE Trans Syst Man

Cybern B Cybern, 2008. 38(3): p. 771-84.

Plewczynski, D., S. A. Spieser, and U. Koch, Performance

of machine learning methods for ligand-based virtual

screening. Comb Chem High Throughput Screen,

2009. 12(4): p. 358-68.

Schwenk, H. and Y. Bengio, Boosting neural networks.

Neural Comput, 2000. 12(8): p. 1869-87.

Serpen, G., D. K. Tekkedil, and M. Orra, A knowledge-

based artificial neural network classifier for

pulmonary embolism diagnosis. Comput Biol Med,

2008. 38(2): p. 204-20.

Shrestha, D. L. and D. P. Solomatine, Experiments with

AdaBoost.RT, an improved boosting scheme for

regression. Neural Comput, 2006. 18(7): p. 1678-710.

27. Peng, Y., A novel ensemble machine learning for

robust microarray data classification. Comput Biol

Med, 2006. 36(6): p. 553-73.

Wang, C. W., New ensemble machine learning method for

classification and prediction on gene expression data.

Conf Proc IEEE Eng Med Biol Soc, 2006. 1: p. 3478-

81.

Yang, J. Y., et al., A hybrid machine learning-based

method for classifying the Cushing's Syndrome with

BRAINSTORMING - Agent based Meta-learning Approach

489

comorbid adrenocortical lesions. BMC Genomics,

2008. 9 Suppl 1: p. S23.

Devillers, J., et al., Internet resources for agent-based

modelling. SAR QSAR Environ Res. 21(3-4): p. 337-

50.

Moore, D., et al., Extending drug ethno-epidemiology

using agent-based modelling. Addiction, 2009.

104(12): p. 1991-7.

Farmer, J. D. and D. Foley, The economy needs agent-

based modelling. Nature, 2009. 460(7256): p. 685-6.

Gu, W. and R.J. Novak, Agent-based modelling of

mosquito foraging behaviour for malaria control.

Trans R Soc Trop Med Hyg, 2009. 103(11): p. 1105-

12.

Pogson, M., et al., Introducing spatial information into

predictive NF-kappaB modelling--an agent-based

approach. PLoS ONE, 2008. 3(6): p. e2367.

35. Sun, T., et al., Agent based modelling helps in

understanding the rules by which fibroblasts support

keratinocyte colony formation. PLoS ONE, 2008. 3(5):

p. e2129.

Bryson, J. J., Y. Ando, and H. Lehmann, Agent-based

modelling as scientific method: a case study analysing

primate social behaviour. Philos Trans R Soc Lond B

Biol Sci, 2007. 362(1485): p. 1685-98.

Pogson, M., et al., Formal agent-based modelling of

intracellular chemical interactions. Biosystems, 2006.

85(1): p. 37-45.

Walker, D. C., et al., The epitheliome: agent-based

modelling of the social behaviour of cells. Biosystems,

2004. 76(1-3): p. 89-100.

Plewczynski, D., Mean-field theory of meta-learning.

Journal of Statistical Mechanics-Theory and

Experiment, 2009: p. -.

Plewczynski, D., Landau theory of social clustering.

Physica A, 1998. 261(3-4): p. 608-617.

Plewczynski, D., Brainstorming: weighted voting

prediction of inhibitors for protein targets. J Mol

Model, 2010.

ICAART 2011 - 3rd International Conference on Agents and Artificial Intelligence

490