Douglas Adriano Augusto

Symbolic regression via genetic programming

Presents an implementation of symbolic regression which is based on genetic programming (GP). Unfortunately, standard implementations of GP in compiled languages are not usually the most efficient ones. The present approach employs a simple representation for tree-like structures by making use of Reads linear code, leading to more simplicity and better performance when compared with traditional GP implementations. Creation, crossover and mutation of individuals are formalized. An extension allowing for the creation of random coefficients is presented. The efficiency of the proposed implementation was confirmed in computational experiments which are summarized in the paper.

An immune-inspired instance selection mechanism for supervised classification

One issue in data classification problems is to find an optimal subset of instances to train a classifier. Training sets that represent well the characteristics of each class have better chances to build a successful predictor. There are cases where data are redundant or take large amounts of computing time in the learning process. To overcome this issue, instance selection techniques have been proposed. These techniques remove examples from the data set so that classifiers are built faster and, in some cases, with better accuracy. Some of these techniques are based on nearest neighbors, ordered removal, random sampling and evolutionary methods. The weaknesses of these methods generally involve lack of accuracy, overfitting, lack of robustness when the data set size increases and high complexity. This work proposes a simple and fast immune-inspired suppressive algorithm for instance selection, called SeleSup. According to self-regulation mechanisms, those cells unable to neutralize danger tend to disappear from the organism. Therefore, by analogy, data not relevant to the learning of a classifier are eliminated from the training process. The proposed method was compared with three important instance selection algorithms on a number of data sets. The experiments showed that our mechanism substantially reduces the data set size and is accurate and robust, specially on larger data sets.

NICeSim: An open-source simulator based on machine learning techniques to support medical research on prenatal and perinatal care decision making

OBJECTIVE This paper describes NICeSim, an open-source simulator that uses machine learning (ML) techniques to aid health professionals to better understand the treatment and prognosis of premature newborns. METHODS The application was developed and tested using data collected in a Brazilian hospital. The available data were used to feed an ML pipeline that was designed to create a simulator capable of predicting the outcome (death probability) for newborns admitted to neonatal intensive care units. However, unlike previous scoring systems, our computational tool is not intended to be used at the patients bedside, although it is possible. Our primary goal is to deliver a computational system to aid medical research in understanding the correlation of key variables with the studied outcome so that new standards can be established for future clinical decisions. In the implemented simulation environment, the values of key attributes can be changed using a user-friendly interface, where the impact of each change on the outcome is immediately reported, allowing a quantitative analysis, in addition to a qualitative investigation, and delivering a totally interactive computational tool that facilitates hypothesis construction and testing. RESULTS Our statistical experiments showed that the resulting model for death prediction could achieve an accuracy of 86.7% and an area under the receiver operating characteristic curve of 0.84 for the positive class. Using this model, three physicians and a neonatal nutritionist performed simulations with key variables correlated with chance of death. The results indicated important tendencies for the effect of each variable and the combination of variables on prognosis. We could also observe values of gestational age and birth weight for which a low Apgar score and the occurrence of respiratory distress syndrome (RDS) could be less or more severe. For instance, we have noticed that for a newborn with 2000 g or more the occurrence of RDS is far less problematic than for neonates weighing less. CONCLUSIONS The significant accuracy demonstrated by our predictive model shows that NICeSim might be used for hypothesis testing to minimize in vivo experiments. We observed that the model delivers predictions that are in very good agreement with the literature, demonstrating that NICeSim might be an important tool for supporting decision making in medical practice. Other very important characteristics of NICeSim are its flexibility and dynamism. NICeSim is flexible because it allows the inclusion and deletion of variables according to the requirements of a particular study. It is also dynamic because it trains a just-in-time model. Therefore, the system is improved as data from new patients become available. Finally, NICeSim can be extended in a cooperative manner because it is an open-source system.

Coevolutionary multi-population genetic programming for data classification

This work presents a new evolutionary ensemble method for data classification, which is inspired by the concepts of bagging and boosting, and aims at combining their good features while avoiding their weaknesses. The approach is based on a distributed multiple-population genetic programming (GP) algorithm which exploits the technique of coevolution at two levels. On the inter-population level the populations cooperate in a semi-isolated fashion, whereas on the intra-population level the candidate classifiers coevolve competitively with the training data samples. The final classifier is a voting committee composed by the best members of all the populations. The experiments performed in a varying number of populations show that our approach outperforms both bagging and boosting for a number of benchmark problems.

Coevolution of data samples and classifiers integrated with grammatically-based genetic programming for data classification

The present work treats the data classification task by means of evolutionary computation techniques using three ingredients: genetic programming, competitive coevolution, and context-free grammar. The robustness and symbolic/interpretative qualities of the genetic programming are employed to construct classification trees via Darwinian evolution. The flexible formal structure of the context-free grammar replaces the standard genetic programming representation and describes a language which encodes trees of varying complexity. Finally, competitive coevolution is used to promote competitions between data samples and classification trees in order to create and sustain an evolutionary arms-race for improved solutions.

Evolving numerical constants in grammatical evolution with the ephemeral constant method

This paper assesses the new numerical-constant generation method called ephemeral constant, which can be seen as a translation of the classical genetic programmings ephemeral random constant to the grammatical evolution framework. Its most distinctive feature is that it decouples the number of bits used to encode the grammars production rules from the number of bits used to represent a constant. This makes it possible to increase the methods representational power without incurring in an overly redundant encoding scheme. We present experiments comparing ephemeral constant with the three most popular approaches for constant handling: the traditional approach, digit concatenation, and persistent random constant. By varying the number of bits to represent a constant, we can increase the numerical precision to the desired level of accuracy, overcoming by a large margin the other approaches.

A new approach for generating numerical constants in grammatical evolution

A new approach for numerical-constant generation in Grammatical Evolution is presented. Experiments comparing our method with the three most popular methods for constant creation are performed. By varying the number of bits to represent a constant, we can increase our methods precision to the desired level of accuracy, overcoming by a large margin the other approaches.

On the characteristics of sequential decision problems and their impact on evolutionary computation and reinforcement learning

This work provides a systematic review of the criteria most commonly used to classify sequential decision problems and discusses their impact on the performance of reinforcement learning and evolutionary computation. The paper also proposes a further division of one class of decision problems into two subcategories, which delimits a set of decision tasks particularly difficult for optimization techniques in general and evolutionary methods in particular. A simple computational experiment is presented to illustrate the subject.

Evaluating the feasibility of grammar-based GP in combining meteorological forecast models

The purpose of this paper is to evaluate the feasibility of grammatical evolution (GE) in combining meteorological models into more accurate single forecast of rainfall amount. A set of GE experiments was performed comparing six proposed ensemble forecast grammars on three benchmark problems. We also proposed a manner of designing benchmark problems by creating arbitrary combinations of meteorological models, as well as modeling the effect of weather patterns over models as explicit functions. The results showed that the GE algorithm obtained a superior performance relative to three traditional statistical methods for all the benchmark problems. A comparison among the developed grammars showed that our most complex grammar, which allows non-linear combinations of models and an unrestricted use of patterns, turned out to be the overall best performing proposal.

Predicting the Performance of Job Applicants by Means of Genetic Programming

Since their early development, genetic programming-based algorithms have been showing to be successful at challenging problems, attaining several human-competitive results and other awards. This paper will present another achievement of such algorithms by describing how our team has won an international machine-learning competition. We have solved, by means of grammar-based genetic programming techniques, a real-world problem of meritocracy in jobs by evolving classifiers that were both accurate and human-readable.