En torno a los efectos de la edad en el aprendizaje escolar de una lengua extrangera

The purpose of this article is to treat a currently much debated issue, the effects of age on second language learning. To do so, we contrast data collected by our research team from over one thousand seven hundred young and adult learners with four popular beliefs or generalizations, which, while deeply rooted in this society, are not always corroborated by our data.Two of these generalizations about Second Language Acquisition (languages spoken in the social context) seem to be widely accepted: a) older children, adolescents and adults are quicker and more efficient at the first stages of learning than are younger learners; b) in a natural context children with an early start are more liable to attain higher levels of proficiency. However, in the context of Foreign Language Acquisition, the context in which we collect the data, this second generalization is difficult to verify due to the low number of instructional hours (a maximum of some 800 hours) and the lower levels of language exposure time provided. The design of our research project has allowed us to study differences observed with respect to the age of onset (ranging from 2 to 18+), but in this article we focus on students who began English instruction at the age of 8 (LOGSE Educational System) and those who began at the age of 11 (EGB). We have collected data from both groups after a period of 200 (Time 1) and 416 instructional hours (Time 2), and we are currently collecting data after a period of 726 instructional hours (Time 3). We have designed and administered a variety of tests: tests on English production and reception, both oral and written, and within both academic and communicative oriented approaches, on the learners' L1 (Spanish and Catalan), as well as a questionnaire eliciting personal and sociolinguistic information. The questions we address and the relevant empirical evidence are as follows: 1. "For young children, learning languages is a game. They enjoy it more than adults."Our data demonstrate that the situation is not quite so. Firstly, both at the levels of Primary and Secondary education (ranging from 70.5% in 11-year-olds to 89% in 14-year-olds) students have a positive attitude towards learning English. Secondly, there is a difference between the two groups with respect to the factors they cite as responsible for their motivation to learn English: the younger students cite intrinsic factors, such as the games they play, the methodology used and the teacher, whereas the older students cite extrinsic factors, such as the role of their knowledge of English in the achievement of their future professional goals. 2 ."Young children have more resources to learn languages." Here our data suggest just the opposite. The ability to employ learning strategies (actions or steps used) increases with age. Older learners' strategies are more varied and cognitively more complex. In contrast, younger learners depend more on their interlocutor and external resources and therefore have a lower level of autonomy in their learning. 3. "Young children don't talk much but understand a lot"This third generalization does seem to be confirmed, at least to a certain extent, by our data in relation to the analysis of differences due to the age factor and productive use of the target language. As seen above, the comparably slower progress of the younger learners is confirmed. Our analysis of interpersonal receptive abilities demonstrates as well the advantage of the older learners. Nevertheless, with respect to passive receptive activities (for example, simple recognition of words or sentences) no great differences are observed. Statistical analyses suggest that in this test, in contrast to the others analyzed, the dominance of the subjects' L1s (reflecting a cognitive capacity that grows with age) has no significant influence on the learning process. 4. "The sooner they begin, the better their results will be in written language"This is not either completely confirmed in our research. First of all, we perceive that certain compensatory strategies disappear only with age, but not with the number of instructional hours. Secondly, given an identical number of instructional hours, the older subjects obtain better results. With respect to our analysis of data from subjects of the same age (12 years old) but with a different number of instructional hours (200 and 416 respectively, as they began at the ages of 11 and 8), we observe that those who began earlier excel only in the area of lexical fluency. In conclusion, the superior rate of older learners appears to be due to their higher level of cognitive development, a factor which allows them to benefit more from formal or explicit instruction in the school context. Younger learners, however, do not benefit from the quantity and quality of linguistic exposure typical of a natural acquisition context in which they would be allowed to make use of implicit learning abilities. It seems clear, then, that the initiative in this country to begin foreign language instruction earlier will have positive effects only if it occurs in combination with either higher levels of exposure time to the foreign language, or, alternatively, with its use as the language of instruction in other areas of the curriculum.

Species och artikelbruk i finskspråkiga grundskoleelevers inlärarsvenska. En longitudinell undersökning i årskurserna 7–9

Denna doktorsavhandling utreder hur finska grundskolelever använder de svenska substantivens bestämdhetsformer och artiklar och hur deras kunskaper utvecklas under årskurserna 7-9. Species och artikelbruk är problematiska för alla andraspråksinlärare i svenska, men de är synnerligen svåra för inlärare vars förstaspråk saknar morfologisk species. Det svenska systemet avviker också kraftigt från det motsvarande systemet i engelskan, varför tidigare kunskaper i engelska inte är till någon stor hjälp i inlärningen, låt vara att bestämdheten som begrepp redan är bekant för inläraren. Den teoretiska referensramen bygger på både grammatiska beskrivningar av den svenska grammatiken och på teorierna om grammatikinlärningen i andraspråk. Bland de sistnämnda är teorierna om tvärspråkligt inflytande, explicit respektive implicit inlärning samt helsekvensinlärning (på engelska formulaic language) av relevans. Undersökningsmaterialet består av korta texter samt inspelat muntligt material som med jämna mellanrum insamlats av finskspråkiga grundskolelever (n=67) som läser B-svenska. Undersökningen är i första hand kvantitativ, om än inmatningen av nominalfraserna i materialet samt deras formella och semantiska aspekter i analysprogrammet Microsoft Access också innebar en omfattande kvalitativ analys. Undersökningen bygger på performansanalysen och analysen av obligatoriska kontexter och beräkningen av frekvenser och korrekthetsprocent för de olika nominalfrastyperna. Informanterna använder komplext språk redan i årskurs 7. Korrekthetsprocenten stiger under undersökningstiden i de flesta frastyperna, men skillnaderna är sällan statistiskt signifikanta. Den normativa analysen visar också, att formfelen är i både det skriftliga och det muntliga materialet signifikant vanligare än speciesfelen. Det är med andra ord lättare för informanterna att välja rätt species än att bilda en korrekt nominalfras. I tidigare undersökningar i Sverige har likadana resultat nåtts. De mest centrala frastyperna i undersökningen bildar i båda typerna av materialet en inlärningsgång som upprepas i alla årskurser och kan förklaras med komplexitetsskillnaderna mellan de olika frastyperna. Informanterna behärskar bäst de frastyper, som varken innehåller artiklar eller ändelser. Näst bäst behärskar de substantivets bestämda form singularis och svagast obestämd form singularis, vars artikel är en klassisk svårighetskälla för finska svenskinlärare. Analysen av informanternas läromedel visar dock att den typiska undervisningsordningen i läromedlen inte motsvarar inlärningsgången som upptäckts i denna undersökning.

A Avaliação na sala de aula de matemática: práticas e concepções

This research aims to understand the assessment practices used by teachers at a public state school in the city of Cunha, Sao Paulo. To this end, we interviewed five mathematics teachers, who answered a questionnaire with five questions. The responses were analyzed according to the rigor of phenomenological research. To understand the investigation region, that is to say, the meaning of evaluation, we proceeded to a review of studies on the subject in authors like Buriasco (2002), Pavanello (2006), Hoffmann (1994), expressive in Mathematics Education that allows us to explain the concept of prevailing interpretation in the area. The phenomenological analysis enabled the development of three categories open revealing the concept of evaluation of teachers investigated. The first shows the review As a way to measure the knowledge acquired by the student. His interpretation leads us to understand that for some teachers, the research subjects, the assessment becomes a method to ' measure ' the knowledge acquired by the student. The second category, expressed by As a way of understanding the student's behavior in class, shows that some of the interviewees understand the evaluation as a medium that reveals and appreciates the ways of the student behave in class. Finally, the third category refers to the evaluation by means of said instruments. On this subject the claim that the assessment is through instruments such that: evidence, exercise lists, among others. In summary, interviews and categories analyzed explain the ways in which the assessment reveals the concept of implicit learning the instruments used in the evaluation practices of teachers interviewed. However, the authors read, evaluation is a necessary and permanent teaching job in teaching, which must follow step by step the process of teaching and learning. It follows, ... (Complete abstract click electronic access below)

A Avaliação na sala de aula de matemática: práticas e concepções

This research aims to understand the assessment practices used by teachers at a public state school in the city of Cunha, Sao Paulo. To this end, we interviewed five mathematics teachers, who answered a questionnaire with five questions. The responses were analyzed according to the rigor of phenomenological research. To understand the investigation region, that is to say, the meaning of evaluation, we proceeded to a review of studies on the subject in authors like Buriasco (2002), Pavanello (2006), Hoffmann (1994), expressive in Mathematics Education that allows us to explain the concept of prevailing interpretation in the area. The phenomenological analysis enabled the development of three categories open revealing the concept of evaluation of teachers investigated. The first shows the review As a way to measure the knowledge acquired by the student. His interpretation leads us to understand that for some teachers, the research subjects, the assessment becomes a method to ' measure ' the knowledge acquired by the student. The second category, expressed by As a way of understanding the student's behavior in class, shows that some of the interviewees understand the evaluation as a medium that reveals and appreciates the ways of the student behave in class. Finally, the third category refers to the evaluation by means of said instruments. On this subject the claim that the assessment is through instruments such that: evidence, exercise lists, among others. In summary, interviews and categories analyzed explain the ways in which the assessment reveals the concept of implicit learning the instruments used in the evaluation practices of teachers interviewed. However, the authors read, evaluation is a necessary and permanent teaching job in teaching, which must follow step by step the process of teaching and learning. It follows, ... (Complete abstract click electronic access below)

The Role of Consciousness in Memory

Conscious events interact with memory systems in learning, rehearsal and retrieval (Ebbinghaus 1885/1964; Tulving 1985). Here we present hypotheses that arise from the IDA computional model (Franklin, Kelemen and McCauley 1998; Franklin 2001b) of global workspace theory (Baars 1988, 2002). Our primary tool for this exploration is a flexible cognitive cycle employed by the IDA computational model and hypothesized to be a basic element of human cognitive processing. Since cognitive cycles are hypothesized to occur five to ten times a second and include interaction between conscious contents and several of the memory systems, they provide the means for an exceptionally fine-grained analysis of various cognitive tasks. We apply this tool to the small effect size of subliminal learning compared to supraliminal learning, to process dissociation, to implicit learning, to recognition vs. recall, and to the availability heuristic in recall. The IDA model elucidates the role of consciousness in the updating of perceptual memory, transient episodic memory, and procedural memory. In most cases, memory is hypothesized to interact with conscious events for its normal functioning. The methodology of the paper is unusual in that the hypotheses and explanations presented are derived from an empirically based, but broad and qualitative computational model of human cognition.

Impacts of a word-picture training on literacy skills in elementary school children and youths with intellectual disabilities

Objective: There is convincing evidence that phonological, orthographic and semantic processes influence children’s ability to learn to read and spell words. So far only a few studies investigated the influence of implicit learning in literacy skills. Children are sensitive to the statistics of their learning environment. By frequent reading they acquire implicit knowledge about the frequency of letter patterns in written words, and they use this knowledge during reading and spelling. Additionally, semantic connections facilitate to storing of words in memory. Thus, the aim of the intervention study was to implement a word-picture training which is based on statistical and semantic learning. Furthermore, we aimed at examining the training effects in reading and spelling in comparison to an auditory-visual matching training and a working memory training program. Participants and Methods: One hundred and thirty-two children aged between 8 and 11 years participated in training in three weekly session of 12 minutes over 8 weeks, and completed other assessments of reading, spelling, working memory and intelligence before and after training. Results: Results revealed in general that the word-picture training and the auditory-visual matching training led to substantial gains in reading and spelling performance in comparison to the working-memory training. Although both children with and without learning difficulties profited in their reading and spelling after the word-picture training, the training program led to differential effects for the two groups. After the word-picture training on the one hand, children with learning difficulties profited more in spelling as children without learning difficulties, on the other hand, children without learning difficulties benefit more in word comprehension. Conclusions: These findings highlight the need for frequent reading trainings with semantic connections in order to support the acquisition of literacy skills.

Impacts of a word-picture training on reading in youth with mixed intellectual disabilities. A waiting-list control group comparison

Individuals with intellectual disabilities (ID) often struggle with learning how to read. Reading difficulties seem to be the most common secondary condition of ID. Only one in five children with mild or moderate ID achieves even minimal literacy skills. However, literacy education for children and adolescents with ID has been largely overlooked by researchers and educators. While there is little research on reading of children with ID, many training studies have been conducted with other populations with reading difficulties. The most common approach of acquiring literacy skills consists of sophisticated programs that train phonological skills and auditory perception. Only few studies investigated the influence of implicit learning on literacy skills. Implicit learning processes seem to be largely independent of age and IQ. Children are sensitive to the statistics of their learning environment. By frequent word reading they acquire implicit knowledge about the frequency of single letters and letter patterns in written words. Additionally, semantic connections not only improve the word understanding, but also facilitate storage of words in memory. Advances in communication technology have introduced new possibilities for remediating literacy skills. Computers can provide training material in attractive ways, for example through animations and immediate feedback .These opportunities can scaffold and support attention processes central to learning. Thus, the aim of this intervention study was to develop and implement a computer based word-picture training, which is based on statistical and semantic learning, and to examine the training effects on reading, spelling and attention in children and adolescents (9-16 years) diagnosed with mental retardation (general IQ  74). Fifty children participated in four to five weekly training sessions of 15-20 minutes over 4 weeks, and completed assessments of attention, reading, spelling, short-term memory and fluid intelligence before and after training. After a first assessment (T1), the entire sample was divided in a training group (group A) and a waiting control group (group B). After 4 weeks of training with group A, a second assessment (T2) was administered with both training groups. Afterwards, group B was trained for 4 weeks, before a last assessment (T3) was carried out in both groups. Overall, the results showed that the word-picture training led to substantial gains on word decoding and attention for both training groups. These effects were preserved six weeks later (group A). There was also a clear tendency of improvement in spelling after training for both groups, although the effect did not reach significance. These findings highlight the fact that an implicit statistical learning training in a playful way by motivating computer programs can not only promote reading development, but also attention in children with intellectual disabilities.

The impact of the position effect on the factorial structure of the Culture Fair Test (CFT)

The Culture Fair Test (CFT) is a psychometric test of fluid intelligence consisting of four subtests; Series, Classification, Matrices, and Topographies. The four subtests are only moderately intercorrelated, doubting the notion that they assess the same construct (i.e., fluid intelligence). As an explanation of these low correlations, we investigated the position effect. This effect is assumed to reflect implicit learning during testing. By applying fixed-links modeling to analyze the CFT data of 206 participants, we identified position effects as latent variables in the subtests; Classification, Matrices, and Topographies. These position effects were disentangled from a second set of latent variables representing fluid intelligence inherent in the four subtests. After this separation of position effect and basic fluid intelligence, the latent variables representing basic fluid intelligence in the subtests Series, Matrices, and Topographies could be combined to one common latent variable which was highly correlated with fluid intelligence derived from the subtest Classification (r=.72). Correlations between the three latent variables representing the position effects in the Classification, Matrices, and Topographies subtests ranged from r=.38 to r=.59. The results indicate that all four CFT subtests measure the same construct (i.e., fluid intelligence) but that the position effect confounds the factorial structure

The position effect in a Rasch-homogenous test: A fixed-links modeling approach.

The position effect describes the influence of just-completed items in a psychological scale on subsequent items. This effect has been repeatedly reported for psychometric reasoning scales and is assumed to reflect implicit learning during testing. One way to identify the position effect is fixed-links modeling. With this approach, two latent variables are derived from the test items. Factor loadings of one latent variable are fixed to 1 for all items to represent ability-related variance. Factor loadings on the second latent variable increase from the first to the last item describing the position effect. Previous studies using fixed-links modeling on the position effect investigated reasoning scales constructed in accordance with classical test theory (e.g., Raven’s Progressive Matrices) but, to the best of our knowledge, no Rasch-scaled tests. These tests, however, meet stronger requirements on item homogeneity. In the present study, therefore, we will analyze data from 239 participants who have completed the Rasch-scaled Viennese Matrices Test (VMT). Applying a fixed-links modeling approach, we will test whether a position effect can be depicted as a latent variable and separated from a latent variable representing basic reasoning ability. The results have implications for the assumption of homogeneity in Rasch-homogeneous tests.

Toward a molecular definition of long-term memory storage

The storage of long-term memory is associated with a cellular program of gene expression, altered protein synthesis, and the growth of new synaptic connections. Recent studies of a variety of memory processes, ranging in complexity from those produced by simple forms of implicit learning in invertebrates to those produced by more complex forms of explicit learning in mammals, suggest that part of the molecular switch required for consolidation of long-term memory is the activation of a cAMP-inducible cascade of genes and the recruitment of cAMP response element binding protein-related transcription factors. This conservation of steps in the mechanisms for learning-related synaptic plasticity suggests the possibility of a molecular biology of cognition.

'A Bayesian Optimisation Algorithm for the Nurse Scheduling Problem'

Abstract- A Bayesian optimization algorithm for the nurse scheduling problem is presented, which involves choosing a suitable scheduling rule from a set for each nurse's assignment. Unlike our previous work that used GAs to implement implicit learning, the learning in the proposed algorithm is explicit, i.e. eventually, we will be able to identify and mix building blocks directly. The Bayesian optimization algorithm is applied to implement such explicit learning by building a Bayesian network of the joint distribution of solutions. The conditional probability of each variable in the network is computed according to an initial set of promising solutions. Subsequently, each new instance for each variable is generated by using the corresponding conditional probabilities, until all variables have been generated, i.e. in our case, a new rule string has been obtained. Another set of rule strings will be generated in this way, some of which will replace previous strings based on fitness selection. If stopping conditions are not met, the conditional probabilities for all nodes in the Bayesian network are updated again using the current set of promising rule strings. Computational results from 52 real data instances demonstrate the success of this approach. It is also suggested that the learning mechanism in the proposed approach might be suitable for other scheduling problems.

'Bayesian Optimisation for Nurse Scheduling'

A Bayesian optimisation algorithm for a nurse scheduling problem is presented, which involves choosing a suitable scheduling rule from a set for each nurse's assignment. When a human scheduler works, he normally builds a schedule systematically following a set of rules. After much practice, the scheduler gradually masters the knowledge of which solution parts go well with others. He can identify good parts and is aware of the solution quality even if the scheduling process is not yet completed, thus having the ability to finish a schedule by using flexible, rather than fixed, rules. In this paper, we design a more human-like scheduling algorithm, by using a Bayesian optimisation algorithm to implement explicit learning from past solutions. A nurse scheduling problem from a UK hospital is used for testing. Unlike our previous work that used Genetic Algorithms to implement implicit learning [1], the learning in the proposed algorithm is explicit, i.e. we identify and mix building blocks directly. The Bayesian optimisation algorithm is applied to implement such explicit learning by building a Bayesian network of the joint distribution of solutions. The conditional probability of each variable in the network is computed according to an initial set of promising solutions. Subsequently, each new instance for each variable is generated by using the corresponding conditional probabilities, until all variables have been generated, i.e. in our case, new rule strings have been obtained. Sets of rule strings are generated in this way, some of which will replace previous strings based on fitness. If stopping conditions are not met, the conditional probabilities for all nodes in the Bayesian network are updated again using the current set of promising rule strings. For clarity, consider the following toy example of scheduling five nurses with two rules (1: random allocation, 2: allocate nurse to low-cost shifts). In the beginning of the search, the probabilities of choosing rule 1 or 2 for each nurse is equal, i.e. 50%. After a few iterations, due to the selection pressure and reinforcement learning, we experience two solution pathways: Because pure low-cost or random allocation produces low quality solutions, either rule 1 is used for the first 2-3 nurses and rule 2 on remainder or vice versa. In essence, Bayesian network learns 'use rule 2 after 2-3x using rule 1' or vice versa. It should be noted that for our and most other scheduling problems, the structure of the network model is known and all variables are fully observed. In this case, the goal of learning is to find the rule values that maximize the likelihood of the training data. Thus, learning can amount to 'counting' in the case of multinomial distributions. For our problem, we use our rules: Random, Cheapest Cost, Best Cover and Balance of Cost and Cover. In more detail, the steps of our Bayesian optimisation algorithm for nurse scheduling are: 1. Set t = 0, and generate an initial population P(0) at random; 2. Use roulette-wheel selection to choose a set of promising rule strings S(t) from P(t); 3. Compute conditional probabilities of each node according to this set of promising solutions; 4. Assign each nurse using roulette-wheel selection based on the rules' conditional probabilities. A set of new rule strings O(t) will be generated in this way; 5. Create a new population P(t+1) by replacing some rule strings from P(t) with O(t), and set t = t+1; 6. If the termination conditions are not met (we use 2000 generations), go to step 2. Computational results from 52 real data instances demonstrate the success of this approach. They also suggest that the learning mechanism in the proposed approach might be suitable for other scheduling problems. Another direction for further research is to see if there is a good constructing sequence for individual data instances, given a fixed nurse scheduling order. If so, the good patterns could be recognized and then extracted as new domain knowledge. Thus, by using this extracted knowledge, we can assign specific rules to the corresponding nurses beforehand, and only schedule the remaining nurses with all available rules, making it possible to reduce the solution space. Acknowledgements The work was funded by the UK Government's major funding agency, Engineering and Physical Sciences Research Council (EPSRC), under grand GR/R92899/01. References [1] Aickelin U, "An Indirect Genetic Algorithm for Set Covering Problems", Journal of the Operational Research Society, 53(10): 1118-1126,

An Estimation of Distribution Algorithm for Nurse Scheduling

Schedules can be built in a similar way to a human scheduler by using a set of rules that involve domain knowledge. This paper presents an Estimation of Distribution Algorithm (EDA) for the nurse scheduling problem, which involves choosing a suitable scheduling rule from a set for the assignment of each nurse. Unlike previous work that used Genetic Algorithms (GAs) to implement implicit learning, the learning in the proposed algorithm is explicit, i.e. we identify and mix building blocks directly. The EDA is applied to implement such explicit learning by building a Bayesian network of the joint distribution of solutions. The conditional probability of each variable in the network is computed according to an initial set of promising solutions. Subsequently, each new instance for each variable is generated by using the corresponding conditional probabilities, until all variables have been generated, i.e. in our case, a new rule string has been obtained. Another set of rule strings will be generated in this way, some of which will replace previous strings based on fitness selection. If stopping conditions are not met, the conditional probabilities for all nodes in the Bayesian network are updated again using the current set of promising rule strings. Computational results from 52 real data instances demonstrate the success of this approach. It is also suggested that the learning mechanism in the proposed approach might be suitable for other scheduling problems.

An Estimation of Distribution Algorithm for Nurse Scheduling

Schedules can be built in a similar way to a human scheduler by using a set of rules that involve domain knowledge. This paper presents an Estimation of Distribution Algorithm (EDA) for the nurse scheduling problem, which involves choosing a suitable scheduling rule from a set for the assignment of each nurse. Unlike previous work that used Genetic Algorithms (GAs) to implement implicit learning, the learning in the proposed algorithm is explicit, i.e. we identify and mix building blocks directly. The EDA is applied to implement such explicit learning by building a Bayesian network of the joint distribution of solutions. The conditional probability of each variable in the network is computed according to an initial set of promising solutions. Subsequently, each new instance for each variable is generated by using the corresponding conditional probabilities, until all variables have been generated, i.e. in our case, a new rule string has been obtained. Another set of rule strings will be generated in this way, some of which will replace previous strings based on fitness selection. If stopping conditions are not met, the conditional probabilities for all nodes in the Bayesian network are updated again using the current set of promising rule strings. Computational results from 52 real data instances demonstrate the success of this approach. It is also suggested that the learning mechanism in the proposed approach might be suitable for other scheduling problems.

'Bayesian Optimisation for Nurse Scheduling'

A Bayesian optimisation algorithm for a nurse scheduling problem is presented, which involves choosing a suitable scheduling rule from a set for each nurse's assignment. When a human scheduler works, he normally builds a schedule systematically following a set of rules. After much practice, the scheduler gradually masters the knowledge of which solution parts go well with others. He can identify good parts and is aware of the solution quality even if the scheduling process is not yet completed, thus having the ability to finish a schedule by using flexible, rather than fixed, rules. In this paper, we design a more human-like scheduling algorithm, by using a Bayesian optimisation algorithm to implement explicit learning from past solutions. A nurse scheduling problem from a UK hospital is used for testing. Unlike our previous work that used Genetic Algorithms to implement implicit learning [1], the learning in the proposed algorithm is explicit, i.e. we identify and mix building blocks directly. The Bayesian optimisation algorithm is applied to implement such explicit learning by building a Bayesian network of the joint distribution of solutions. The conditional probability of each variable in the network is computed according to an initial set of promising solutions. Subsequently, each new instance for each variable is generated by using the corresponding conditional probabilities, until all variables have been generated, i.e. in our case, new rule strings have been obtained. Sets of rule strings are generated in this way, some of which will replace previous strings based on fitness. If stopping conditions are not met, the conditional probabilities for all nodes in the Bayesian network are updated again using the current set of promising rule strings. For clarity, consider the following toy example of scheduling five nurses with two rules (1: random allocation, 2: allocate nurse to low-cost shifts). In the beginning of the search, the probabilities of choosing rule 1 or 2 for each nurse is equal, i.e. 50%. After a few iterations, due to the selection pressure and reinforcement learning, we experience two solution pathways: Because pure low-cost or random allocation produces low quality solutions, either rule 1 is used for the first 2-3 nurses and rule 2 on remainder or vice versa. In essence, Bayesian network learns 'use rule 2 after 2-3x using rule 1' or vice versa. It should be noted that for our and most other scheduling problems, the structure of the network model is known and all variables are fully observed. In this case, the goal of learning is to find the rule values that maximize the likelihood of the training data. Thus, learning can amount to 'counting' in the case of multinomial distributions. For our problem, we use our rules: Random, Cheapest Cost, Best Cover and Balance of Cost and Cover. In more detail, the steps of our Bayesian optimisation algorithm for nurse scheduling are: 1. Set t = 0, and generate an initial population P(0) at random; 2. Use roulette-wheel selection to choose a set of promising rule strings S(t) from P(t); 3. Compute conditional probabilities of each node according to this set of promising solutions; 4. Assign each nurse using roulette-wheel selection based on the rules' conditional probabilities. A set of new rule strings O(t) will be generated in this way; 5. Create a new population P(t+1) by replacing some rule strings from P(t) with O(t), and set t = t+1; 6. If the termination conditions are not met (we use 2000 generations), go to step 2. Computational results from 52 real data instances demonstrate the success of this approach. They also suggest that the learning mechanism in the proposed approach might be suitable for other scheduling problems. Another direction for further research is to see if there is a good constructing sequence for individual data instances, given a fixed nurse scheduling order. If so, the good patterns could be recognized and then extracted as new domain knowledge. Thus, by using this extracted knowledge, we can assign specific rules to the corresponding nurses beforehand, and only schedule the remaining nurses with all available rules, making it possible to reduce the solution space. Acknowledgements The work was funded by the UK Government's major funding agency, Engineering and Physical Sciences Research Council (EPSRC), under grand GR/R92899/01. References [1] Aickelin U, "An Indirect Genetic Algorithm for Set Covering Problems", Journal of the Operational Research Society, 53(10): 1118-1126,