Implicit Statistical Learning of Brazilian Children With Developmental Dyslexia.

Implicit statistical learning (ISL) is non-conscious learning where participants identify patterns in their environment after repeated exposures. This study verified whether Brazilian children with DD present disturbances in the ISL mechanism and if these disturbances may be related to the difficulties associated with DD through artificial grammar learning (AGL) and reaction time in serial tasks (SRT Task). It also intended to verify which of the paradigms proves to be the most sensitive to assess ISL and which is most associated with participants' learning to read and write. Two groups of children with and without DD from the end of the first cycle and the beginning of the second cycle of the elementary school participated in this study, paired according to socioeconomic level, education network, schooling, gender and age. Children with DD showed no disturbances in the ISL process; the AGL paradigm exhibited the most significant association with performance on reading/writing tasks. When compared to the SRT Task, the AGL paradigm proved to be more sensitive in assessing implicit processes and effectively distinguishing between the groups with and without DD. The results of the SRT Task emphasise the importance of task practice and structure for implicit learning in children with developmental dyslexia. These findings have important implications for understanding ISL and its relevance to reading and writing skills in children with developmental dyslexia.

Combining statistics: the role of phonotactics on cross-situational word learning

Abstract Language learners can rely on phonological and semantic information to learn novel words. Using a cross-situational word learning paradigm, we explored the role of phonotactic probabilities on word learning in ambiguous contexts. Brazilian-Portuguese speaking adults (N = 30) were exposed to two sets of word-object pairs. Words from one set of labels had slightly higher phonotactic probabilities than words from the other set. By tracking co-occurrences of words and objects, participants were able to learn word-object mappings similarly across both sets. Our findings contrast with studies showing a facilitative effect of phonotactic probability on word learning in non-ambiguous contexts.

Premio de Investigación SCHOT 2020: desarrollo y validación de un modelo multivariables de predicción de estadía hospitalaria en pacientes mayores de 65 años sometidos artroplastia total de cadera electiva en Chile utilizando aprendizaje de máquinas / 2020 SCHOT Research award: Development an a multivariable prediction model of hospital stay in elderly chilean patients undergoing elective total hip arthroplasty using machine learning

INTRODUCCIÓN La predicción de la estadía hospitalaria luego de una artroplastia total de cadera (ATC) electiva es crucial en la evaluación perioperatoria de los pacientes, con un rol determinante desde el punto de vista operacional y económico. Internacionalmente, se han empleado macrodatos (big data, en inglés) e inteligencia artificial para llevar a cabo evaluaciones pronósticas de este tipo. El objetivo del presente estudio es desarrollar y validar, con el empleo del aprendizaje de máquinas (machine learning, en inglés), una herramienta capaz de predecir la estadía hospitalaria de pacientes chilenos mayores de 65 años sometidos a ATC por artrosis. MATERIALES Y MÉTODOS Empleando los registros electrónicos de egresos hospitalarios anonimizados del Departamento de Estadísticas e Información de Salud (DEIS), se obtuvieron los datos de 8.970 egresos hospitalarios de pacientes sometidos a ATC por artrosis entre los años 2016 y 2018. En total, 15 variables disponibles en el DEIS, además del porcentaje de pobreza de la comuna de origen del paciente, fueron incluidos para predecir la probabilidad de que un paciente presentara una estadía acortada (< 3 días) o prolongada (> 3 días) luego de la cirugía. Utilizando técnicas de aprendizaje de máquinas, 8 algoritmos de predicción fueron entrenados con el 80% de la muestra. El 20% restante se empleó para validar las capacidades predictivas de los modelos creados a partir de los algoritmos. La métrica de optimización se evaluó y ordenó en un ranking utilizando el área bajo la curva de característica operativa del receptor (area under the receiver operating characteristic curve, AUC-ROC, en inglés), que corresponde a cuan bien un modelo puede distinguir entre dos grupos. RESULTADOS El algoritmo XGBoost obtuvo el mejor desempeño, con una AUC-ROC promedio de 0,86 (desviación estándar [DE]: 0,0087). En segundo lugar, observamos que el algoritmo lineal de máquina de vector de soporte (support vector machine, SVM, en inglés) obtuvo una AUC-ROC de 0,85 (DE: 0,0086). La importancia relativa de las variables explicativas demostró que la región de residencia, el servicio de salud, el establecimiento de salud donde se operó el paciente, y la modalidad de atención son las variables que más determinan el tiempo de estadía de un paciente. DISCUSIÓN El presente estudio desarrolló algoritmos de aprendizaje de máquinas basados en macrodatos chilenos de libre acceso, y logró desarrollar y validar una herramienta que demuestra una adecuada capacidad discriminatoria para predecir la probabilidad de estadía hospitalaria acortada versus prolongada en adultos mayores sometidos a ATC por artrosis. CONCLUSIÓN Los algoritmos creados a traves del empleo del aprendizaje de máquinas permiten predecir la estadía hospitalaria en pacientes chilenos operado de artroplastia total de cadera electiva

Introduction The prediction of the length of hospital stay after elective total hip arthroplasty (THA) is crucial in the perioperative evaluation of the patients, and it plays a decisive role from the operational and economic point of view. Internationally, big data and artificial intelligence have been used to perform prognostic evaluations of this type. The present study aims to develop and validate, through the use of artificial intelligence (machine learning), a tool capable of predicting the hospital stay of patients over 65 years of age undergoing THA for osteoarthritis. Material and Methods Using the electronic records of hospital discharges de-identified from the Department of Health Statistics and Information (Departamento de Estadísticas e Información de Salud, DEIS, in Spanish), the data of 8,970 hospital discharges of patients who had undergone THA for osteoarthritis between 2016 and 2018 were obtained. A total of 15 variables available in the DEIS registry, in addition to the poverty rate in the patient's borough of origin were included to predict the probability that a patient would have a shortened (< 3 days) or prolonged (> 3 days) stay after surgery. By using machine learning techniques, 8 prediction algorithms were trained with 80% of the sample. The remaining 20% was used to validate the predictive capabilities of the models created from the algorithms. The optimization metric was evaluated and ranked using the area under the receiver operating characteristic curve (AUC-ROC), which corresponds to how well a model can distinguish between two groups. Results The XGBoost algorithm had the best performance, with an average AUC-ROC of 0.86 (standard deviation [SD]: 0.0087). Secondly, we observed that the linear support vector machine (SVM) algorithm obtained an AUC-ROC of 0.85 (SD: 0.0086). The relative importance of the explanatory variables showed that the region of residence, the administrative health service, the hospital where the patient was operated on, and the care modality are the variables that most determine the length of stay. Discussion The present study developed machine learning algorithms based on freeaccess Chilean big data, which helped create and validate a tool that demonstrates an adequate discriminatory capacity to predict shortened versus prolonged hospital stay in elderly patients undergoing elective THA. Conclusion The algorithms created through the use of machine learning allow to predict the hospital stay in Chilean patients undergoing elective total hip arthroplasty Introduction The prediction of the length of hospital stay after elective total hip arthroplasty (THA) is crucial in the perioperative evaluation of the patients, and it plays a decisive role from the operational and economic point of view. Internationally, big data and artificial intelligence have been used to perform prognostic evaluations of this type. The present study aims to develop and validate, through the use of artificial intelligence (machine learning), a tool capable of predicting the hospital stay of patients over 65 years of age undergoing THA for osteoarthritis. Material and Methods Using the electronic records of hospital discharges de-identified from the Department of Health Statistics and Information (Departamento de Estadísticas e Información de Salud, DEIS, in Spanish), the data of 8,970 hospital discharges of patients who had undergone THA for osteoarthritis between 2016 and 2018 were obtained. A total of 15 variables available in the DEIS registry, in addition to the poverty rate in the patient's borough of origin were included to predict the probability that a patient would have a shortened (< 3 days) or prolonged (> 3 days) stay after surgery. By using machine learning techniques, 8 prediction algorithms were trained with 80% of the sample. The remaining 20% was used to validate the predictive capabilities of the models created from the algorithms. The optimization metric was evaluated and ranked using the area under the receiver operating characteristic curve (AUC-ROC), which corresponds to how well a model can distinguish between two groups. Results The XGBoost algorithm had the best performance, with an average AUC-ROC of 0.86 (standard deviation [SD]: 0.0087). Secondly, we observed that the linear

Evidence that haloperidol impairs learning and motivation scores in a probabilistic task by reducing the reward expectation.

Activation of midbrain dopamine neurons in response to positive prediction errors and reward predictive cues is proposed to "energize" reward seeking behaviors and approach responses to places where the reward is expected. In the present study, we tested the effect of the D2-dopamine receptor antagonist haloperidol on response latencies to enter two arms of a Y-maze with different reward probabilities. Adult male Wistar rats were trained to explore the Y-maze with sucrose pellets placed 30% of times at the end of one arm and 70% of times at the opposite arm. Therefore, the reward expectation was different among arms, and was updated in the trials when the reward was omitted. After training, rats received 0.05, 0.10, 0.15 mg/kg haloperidol, or saline 30 min before the test session. In the last, but not in the first trials, haloperidol caused a dose-dependent increase in arm choice latency and response latency. Saline, but not haloperidol, treated rats presented significantly longer response latencies for the 30% compared to the 70% reward probability arm. Haloperidol also caused a dose-dependent decrease in the number of entries in the 70% reward probability arm, increased the number of non-responses, and caused a dose-dependent increase in the number of re-entries in the 30% reward probability arm after non-rewarded trials. Control experiments suggested that haloperidol did not cause motor impairment or satiation, but rather impaired learning and motivation scores by reducing the reward expectation.

Human Foragers: Searchers by Nature and Experience.

Diverse studies of human foraging have revealed behavioral strategies that may have evolved as adaptations for foraging. Here, we used an outdoor experimental search task to explore the effect of three sources of information on participants' performance: (i) information obtained directly from performing a search, (ii) information obtained prior to testing in the form of a distilled snippet of knowledge intended to experimentally simulate information acquired culturally about the environment, and (iii) information obtained from experience of foraging for natural resources for economic gain. We found that (i) immediate searching experience improved performance from the beginning to the end of the short, 2-min task, (ii) information priming improved performance notably from the very beginning of the task, and (iii) natural resource foraging experience improved performance to a lesser extent. Our results highlight the role of culturally transmitted information as well as the presence of mechanisms to rapidly integrate and implement new information into searching choices, which ultimately influence performance in a foraging task.

Statistical learning and spelling: Evidence from Brazilian prephonological spellers.

We analyzed the spelling attempts of Brazilian children (age 3 years, 3 months to 6 years, 0 months) who were prephonological spellers, in that they wrote using letters that did not reflect the phonemes in the words they were asked to spell. We tested the hypothesis that children use their statistical-learning skills to learn about the appearance of writing and that older prephonological spellers, who have had on average more exposure to writing, produce more wordlike spellings than younger prephonological spellers. We found that older prephonological spellers produced longer spellings and were more likely to use letters and digrams in proportion to their frequency of occurrence in Portuguese. There were also some age-related differences in children's tendency to use letters from their own names when writing other words. The results extend previous findings with learners of English to children who are learning a more transparent orthography.

Bayesian validation of grammar productions for the language of thought.

Probabilistic proposals of Language of Thoughts (LoTs) can explain learning across different domains as statistical inference over a compositionally structured hypothesis space. While frameworks may differ on how a LoT may be implemented computationally, they all share the property that they are built from a set of atomic symbols and rules by which these symbols can be combined. In this work we propose an extra validation step for the set of atomic productions defined by the experimenter. It starts by expanding the defined LoT grammar for the cognitive domain with a broader set of arbitrary productions and then uses Bayesian inference to prune the productions from the experimental data. The result allows the researcher to validate that the resulting grammar still matches the intuitive grammar chosen for the domain. We then test this method in the language of geometry, a specific LoT model for geometrical sequence learning. Finally, despite the fact of the geometrical LoT not being a universal (i.e. Turing-complete) language, we show an empirical relation between a sequence's probability and its complexity consistent with the theoretical relationship for universal languages described by Levin's Coding Theorem.

Adaptation in pronoun resolution: Evidence from Brazilian and European Portuguese.

Previous research accounting for pronoun resolution as a problem of probabilistic inference has not explored the phenomenon of adaptation, whereby the processor constantly tracks and adapts, rationally, to changes in a statistical environment. We investigate whether Brazilian (BP) and European Portuguese (EP) speakers adapt to variations in the probability of occurrence of ambiguous overt and null pronouns, in two experiments assessing resolution toward subject and object referents. For each variety (BP, EP), participants were faced with either the same number of null and overt pronouns (equal distribution), or with an environment with fewer overt (than null) pronouns (unequal distribution). We find that the preference for interpreting overt pronouns as referring back to an object referent (object-biased interpretation) is higher when there are fewer overt pronouns (i.e., in the unequal, relative to the equal distribution condition). This is especially the case for BP, a variety with higher prior frequency and smaller object-biased interpretation of overt pronouns, suggesting that participants adapted incrementally and integrated prior statistical knowledge with the knowledge obtained in the experiment. We hypothesize that comprehenders adapted rationally, with the goal of maintaining, across variations in pronoun probability, the likelihood of subject and object referents. Our findings unify insights from research in pronoun resolution and in adaptation, and add to previous studies in both topics: They provide evidence for the influence of pronoun probability in pronoun resolution, and for an adaptation process whereby the language processor not only tracks statistical information, but uses it to make interpretational inferences. (PsycINFO Database Record (c) 2018 APA, all rights reserved).

Exploration and recency as the main proximate causes of probability matching: a reinforcement learning analysis.

Research has not yet reached a consensus on why humans match probabilities instead of maximise in a probability learning task. The most influential explanation is that they search for patterns in the random sequence of outcomes. Other explanations, such as expectation matching, are plausible, but do not consider how reinforcement learning shapes people's choices. We aimed to quantify how human performance in a probability learning task is affected by pattern search and reinforcement learning. We collected behavioural data from 84 young adult participants who performed a probability learning task wherein the majority outcome was rewarded with 0.7 probability, and analysed the data using a reinforcement learning model that searches for patterns. Model simulations indicated that pattern search, exploration, recency (discounting early experiences), and forgetting may impair performance. Our analysis estimated that 85% (95% HDI [76, 94]) of participants searched for patterns and believed that each trial outcome depended on one or two previous ones. The estimated impact of pattern search on performance was, however, only 6%, while those of exploration and recency were 19% and 13% respectively. This suggests that probability matching is caused by uncertainty about how outcomes are generated, which leads to pattern search, exploration, and recency.

Network-based stochastic competitive learning approach to disambiguation in collaborative networks.

Many patterns have been uncovered in complex systems through the application of concepts and methodologies of complex networks. Unfortunately, the validity and accuracy of the unveiled patterns are strongly dependent on the amount of unavoidable noise pervading the data, such as the presence of homonymous individuals in social networks. In the current paper, we investigate the problem of name disambiguation in collaborative networks, a task that plays a fundamental role on a myriad of scientific contexts. In special, we use an unsupervised technique which relies on a particle competition mechanism in a networked environment to detect the clusters. It has been shown that, in this kind of environment, the learning process can be improved because the network representation of data can capture topological features of the input data set. Specifically, in the proposed disambiguating model, a set of particles is randomly spawned into the nodes constituting the network. As time progresses, the particles employ a movement strategy composed of a probabilistic convex mixture of random and preferential walking policies. In the former, the walking rule exclusively depends on the topology of the network and is responsible for the exploratory behavior of the particles. In the latter, the walking rule depends both on the topology and the domination levels that the particles impose on the neighboring nodes. This type of behavior compels the particles to perform a defensive strategy, because it will force them to revisit nodes that are already dominated by them, rather than exploring rival territories. Computer simulations conducted on the networks extracted from the arXiv repository of preprint papers and also from other databases reveal the effectiveness of the model, which turned out to be more accurate than traditional clustering methods.

Predicting sample size required for classification performance.

BACKGROUND: Supervised learning methods need annotated data in order to generate efficient models. Annotated data, however, is a relatively scarce resource and can be expensive to obtain. For both passive and active learning methods, there is a need to estimate the size of the annotated sample required to reach a performance target. METHODS: We designed and implemented a method that fits an inverse power law model to points of a given learning curve created using a small annotated training set. Fitting is carried out using nonlinear weighted least squares optimization. The fitted model is then used to predict the classifier's performance and confidence interval for larger sample sizes. For evaluation, the nonlinear weighted curve fitting method was applied to a set of learning curves generated using clinical text and waveform classification tasks with active and passive sampling methods, and predictions were validated using standard goodness of fit measures. As control we used an un-weighted fitting method. RESULTS: A total of 568 models were fitted and the model predictions were compared with the observed performances. Depending on the data set and sampling method, it took between 80 to 560 annotated samples to achieve mean average and root mean squared error below 0.01. Results also show that our weighted fitting method outperformed the baseline un-weighted method (p < 0.05). CONCLUSIONS: This paper describes a simple and effective sample size prediction algorithm that conducts weighted fitting of learning curves. The algorithm outperformed an un-weighted algorithm described in previous literature. It can help researchers determine annotation sample size for supervised machine learning.

Generality of the summation effect in human causal learning.

Considerable research has examined the contrasting predictions of the elemental and configural association theories proposed by Rescorla and Wagner (1972) and Pearce (1987), respectively. One simple method to distinguish between these approaches is the summation test, in which the associative strength attributed to a novel compound of two separately trained cues is examined. Under common assumptions, the configural view predicts that the strength of the compound will approximate to the average strength of its components, whereas the elemental approach predicts that the strength of the compound will be greater than the strength of either component. Different studies have produced mixed outcomes. In studies of human causal learning, Collins and Shanks (2006) suggested that the observation of summation is encouraged by training, in which different stimuli are associated with different submaximal outcomes, and by testing, in which the alternative outcomes can be scaled. The reported experiments further pursued this reasoning. In Experiment 1, summation was more substantial when the participants were trained with outcomes identified as submaximal than when trained with simple categorical (presence/absence) outcomes. Experiments 2 and 3 demonstrated that summation can also be obtained with categorical outcomes during training, if the participants are encouraged by instruction or the character of training to rate the separately trained components with submaximal ratings. The results are interpreted in terms of apparent performance constraints in evaluations of the contrasting theoretical predictions concerning summation.

Efecto del contenido en la resolución de tareas de conjunción probabilísticas / Content effects in probability judgments

El error de conjunción (Tversky & Kahneman, 1983) se estudió en dos escenarios de probabilidad que suponen distintos contenidos en las tareas a resolver: ficcional y realista (Teigen, Martinussen & Lund, 1996). Participaron voluntariamente 83 sujetos de ambos sexos, alumnos de Psicología de la Universidad de Buenos Aires, quienes resolvieron ambas tareas. Las diferencias halladas en las cantidades de errores de conjunción al comparar las ejecuciones en los dos escenarios fueron altamente significativas. Los resultados reflejan una disminución de los errores cuando se presentan tareas realistas en lugar de ficcionales. Tales hallazgos indican la relevancia de considerar elementos socioecológicos tanto en razonamientos probabilísticos (Hertwig & Gigerenzer, 1999) como en las estrategias didácticas de enseñanza de probabilidad.

Projeção da dinâmica da floresta natural de Terra-firme, região de Manaus-AM, com o uso da cadeia de transição probabilística de Markov / Dynamics of the Terra-firme primary forest in Manaus-AM region using the Markov probabilistic transition matrix

Os fatores que envolvem os processos da dinâmica da floresta influenciam a sua biodiversidade e, portanto, a qualidade da floresta. A definição de estratégias que envolve a proteção e o uso adequado da floresta manejada e a recuperação de áreas já degradadas tornam-se possível com o estudo da estrutura e dinâmica da floresta primária por meio de informações como a mortalidade, o recrutamento e a permanência das árvores no sistema florestal. Este trabalho teve como objetivo avaliar a dinâmica de uma floresta não perturbada e fazer projeções da dinâmica florestal usando a matriz de transição probabilística (Cadeia de Markov). As taxas de recrutamento, mortalidade e incremento foram determinadas a partir de inventários florestais realizados em dois transectos, nos sentidos Norte-Sul e Leste-Oeste (20 x 2500 m cada, totalizando 10 ha), localizados no km 50 da BR 174, na estrada vicinal ZF-2, Manaus/AM, nos anos de 2000 e 2004. A floresta acumulou 8,34 t.ha-1.ano-1 de biomassa fresca acima do solo. De acordo com projeção para 2008, o número total de árvores diminuirá em 2,67 por cento (de 5987 indivíduos (2004) para 5827 (2008)) e a mortalidade será 15 por cento maior (de 264 (2004) para 311 (2008)). O teste Qui-quadrado mostrou que não há diferença significativa (1 por cento de probabilidade) entre as informações coletadas e projetadas. Esses resultados permitem concluir que a Cadeia de Markov é um eficiente instrumento para projetar a dinâmica da floresta natural, contribuindo para o planejamento em curto prazo das atividades que utilizam os recursos florestais.

To combine protection and utilization of forest resources in the tropics, the understanding of forest dynamics is essential. It is also important in the definition of strategies for rehabilitation of degraded areas. In Forestry, forest dynamics could be translated as the understanding of recruitment, mortality and biomass increment rates over time. For this study, these rates were estimated based on measurements carried out in 2000 and 2004 over two transects measuring 20 by 2500 m (5 hectares) each, in Manaus region. This paper deals with forest dynamics of a pristine forest based on the probabilistic transition matrix (the first-order Markov Chain) approach. The main objective is to report 4-year (2000 to 2004) changes in the forest structure. Diameter distribution and tree mortality will be projected ahead to 2008 (t+2), based upon a 4-year period of observations completed in 2004 (t+1) and its immediate past in 2000 (t). In terms of fresh aboveground biomass, this site accumulated 8.34 t.ha-1.ano-1. The chi ² test has shown no statistical difference (p = 0.01) between observed diameter frequency and the expected projected by Markov Chain. This result indicates that the Markov Chain approach is a reliable tool to project the forest dynamics on a short-term basis. In 2008, the total number of individuals will have a decrease of 2.7 percent, and the mortality rate will 15 percent higher than in 2004.

Complex slow potential generators in a simplified attention paradigm.

We have recently obtained evidence for complex multifocal, individually variable generators of slow cortical potentials, elicited during performance of visual tasks involving expecting attention, comparison and memory [Basile, L.F.H., Ballester, G., Castro, C.C., and Gattaz, W.F., 2002. Multifocal slow potential generators revealed by high-resolution EEG and current density reconstruction. Int. J. Psychophysiol., 45 (3), 227-240; Basile, L.F.H, Baldo, M.V., Castro, C.C., and Gattaz, W.F. 2003. The generators of slow potentials obtained during verbal, pictorial and spatial tasks. Int. J. Psychophysiol., 48, 55-65]. The cue-target aspect of traditional paradigms for attention studies is equivalent to 'warning S1'-'imperative S2' in slow potential designs. We simplified Posner's spatial cueing task [Posner, M.I. 1980. Orienting of attention.Q. J. Exp. Psychol. Feb;32 (1), 3-25; Posner, M.I., Snyder, C.R., Davidson, B.J. 1980. Attention and the detection of signals. J Exp Psychol. Jun; 109 (2), 160-174] to temporal cuing only, by using visual cues to indicate the mere presence, on a known central position, of the eventual target (17 ms duration, +/-0.3 degrees grey circle). We recorded slow potentials on 12 healthy subjects, by 124-channel EEG system (Neuroscan Inc.), and modeled their generators using current density reconstruction (CDR) by L(p) 1.2 norm minimization ("Curry V4.6", Neurosoft Inc.) applied to the target onset time. MRIs were obtained for each subject for constraining source models to individual brain anatomy. Average slow potentials were computed from above 60 artifact-free EEG-epochs (ISI=1.6 s, average ITI=2.5 s). We tabulated individual cortical current distributions by cytoarchitectonic area of Brodmann, after scaling into negligible, low, moderate and strong local density, based on percentile bands with respect to absolute maximum current. Despite the task's simplicity, the main result was individual variability and complexity in both scalp voltage and cortical current distributions. As observed in our previous studies, there was strong intersubject variability in the exact distribution of task-related cortical activity. Only parietal area 7 bilaterally was non-negligibly active in all subjects (currents above 10% maximum). As opposed to drawing conclusions based on group averaged data, we propose that activity by cytoarchitectonic area be ranked and statistically analysed only after being scaled on each individual. Based on the present results, the concept of a universal attention-related set of cortical areas if restricted to common areas across subjects is challenged, since even area 7 may no longer be common when the sample size becomes larger. We discuss the fact that group averaging may de-emphasize weakly but consistently active areas, and emphasize strongly but inconsistently active ones.

An adaptive Boolean automation to model circadian cycles.

We propose a Boolean cellular automation to model an artificial adaptive living organism in order to investigate the development of cyclic vital functions during a simulated evolutionary process. The organism is endowed with a basic architecture consisting of several sensor (input), motor (output) and processing Boolean gates whose connectivity pattern is adapted with a genetic algorithm. Cyclic searching behaviors develop that are tuned to the spatial distribution of "food". Under additional assumptions we also find that internal pacemakers can develop to adapt plastically to the alternance of "light" an "darkness". These pacemakers coexist with a "free running" regime in which the circadian cycles persist and even in the absence of external periodic stimuli.