Classificação de nódulos pulmonares em imagens tomográficas utilizando redes neurais artificiais em cascata

Lung cancer is the most common of malignant tumors, with 1.59 million new cases worldwide in 2012. Early detection is the main factor to determine the survival of patients affected by this disease. Furthermore, the correct classification is important to define the most appropriate therapeutic approach as well as suggest the prognosis and the clinical disease evolution. Among the exams used to detect lung cancer, computed tomography have been the most indicated. However, CT images are naturally complex and even experts medical are subject to fault detection or classification. In order to assist the detection of malignant tumors, computer-aided diagnosis systems have been developed to aid reduce the amount of false positives biopsies. In this work it was developed an automatic classification system of pulmonary nodules on CT images by using Artificial Neural Networks. Morphological, texture and intensity attributes were extracted from lung nodules cut tomographic images using elliptical regions of interest that they were subsequently segmented by Otsu method. These features were selected through statistical tests that compare populations (T test of Student and U test of Mann-Whitney); from which it originated a ranking. The features after selected, were inserted in Artificial Neural Networks (backpropagation) to compose two types of classification; one to classify nodules in malignant and benign (network 1); and another to classify two types of malignancies (network 2); featuring a cascade classifier. The best networks were associated and its performance was measured by the area under the ROC curve, where the network 1 and network 2 achieved performance equal to 0.901 and 0.892 respectively.

Bayesian Emulation for Sequential Modeling, Inference and Decision Analysis

The advances in three related areas of state-space modeling, sequential Bayesian learning, and decision analysis are addressed, with the statistical challenges of scalability and associated dynamic sparsity. The key theme that ties the three areas is Bayesian model emulation: solving challenging analysis/computational problems using creative model emulators. This idea defines theoretical and applied advances in non-linear, non-Gaussian state-space modeling, dynamic sparsity, decision analysis and statistical computation, across linked contexts of multivariate time series and dynamic networks studies. Examples and applications in financial time series and portfolio analysis, macroeconomics and internet studies from computational advertising demonstrate the utility of the core methodological innovations.

Chapter 1 summarizes the three areas/problems and the key idea of emulating in those areas. Chapter 2 discusses the sequential analysis of latent threshold models with use of emulating models that allows for analytical filtering to enhance the efficiency of posterior sampling. Chapter 3 examines the emulator model in decision analysis, or the synthetic model, that is equivalent to the loss function in the original minimization problem, and shows its performance in the context of sequential portfolio optimization. Chapter 4 describes the method for modeling the steaming data of counts observed on a large network that relies on emulating the whole, dependent network model by independent, conjugate sub-models customized to each set of flow. Chapter 5 reviews those advances and makes the concluding remarks.

Optimal Capital Requirements in Financial Networks with Fire Sales

I explore and analyze a problem of finding the socially optimal capital requirements for financial institutions considering two distinct channels of contagion: direct exposures among the institutions, as represented by a network and fire sales externalities, which reflect the negative price impact of massive liquidation of assets.These two channels amplify shocks from individual financial institutions to the financial system as a whole and thus increase the risk of joint defaults amongst the interconnected financial institutions; this is often referred to as systemic risk. In the model, there is a trade-off between reducing systemic risk and raising the capital requirements of the financial institutions. The policymaker considers this trade-off and determines the optimal capital requirements for individual financial institutions. I provide a method for finding and analyzing the optimal capital requirements that can be applied to arbitrary network structures and arbitrary distributions of investment returns.

In particular, I first consider a network model consisting only of direct exposures and show that the optimal capital requirements can be found by solving a stochastic linear programming problem. I then extend the analysis to financial networks with default costs and show the optimal capital requirements can be found by solving a stochastic mixed integer programming problem. The computational complexity of this problem poses a challenge, and I develop an iterative algorithm that can be efficiently executed. I show that the iterative algorithm leads to solutions that are nearly optimal by comparing it with lower bounds based on a dual approach. I also show that the iterative algorithm converges to the optimal solution.

Finally, I incorporate fire sales externalities into the model. In particular, I am able to extend the analysis of systemic risk and the optimal capital requirements with a single illiquid asset to a model with multiple illiquid assets. The model with multiple illiquid assets incorporates liquidation rules used by the banks. I provide an optimization formulation whose solution provides the equilibrium payments for a given liquidation rule.

I further show that the socially optimal capital problem using the ``socially optimal liquidation" and prioritized liquidation rules can be formulated as a convex and convex mixed integer problem, respectively. Finally, I illustrate the results of the methodology on numerical examples and

discuss some implications for capital regulation policy and stress testing.

3D considerations for motion perception and visuomotor transformations

Moving through a stable, three-dimensional world is a hallmark of our motor and perceptual experience. This stability is constantly being challenged by movements of the eyes and head, inducing retinal blur and retino-spatial misalignments for which the brain must compensate. To do so, the brain must account for eye and head kinematics to transform two-dimensional retinal input into the reference frame necessary for movement or perception. The four studies in this thesis used both computational and psychophysical approaches to investigate several aspects of this reference frame transformation. In the first study, we examined the neural mechanism underlying the visuomotor transformation for smooth pursuit using a feedforward neural network model. After training, the model performed the general, three-dimensional transformation using gain modulation. This gave mechanistic significance to gain modulation observed in cortical pursuit areas while also providing several testable hypotheses for future electrophysiological work. In the second study, we asked how anticipatory pursuit, which is driven by memorized signals, accounts for eye and head geometry using a novel head-roll updating paradigm. We showed that the velocity memory driving anticipatory smooth pursuit relies on retinal signals, but is updated for the current head orientation. In the third study, we asked how forcing retinal motion to undergo a reference frame transformation influences perceptual decision making. We found that simply rolling one's head impairs perceptual decision making in a way captured by stochastic reference frame transformations. In the final study, we asked how torsional shifts of the retinal projection occurring with almost every eye movement influence orientation perception across saccades. We found a pre-saccadic, predictive remapping consistent with maintaining a purely retinal (but spatially inaccurate) orientation perception throughout the movement. Together these studies suggest that, despite their spatial inaccuracy, retinal signals play a surprisingly large role in our seamless visual experience. This work therefore represents a significant advance in our understanding of how the brain performs one of its most fundamental functions.

Implementação de classificador de tarefas mentais baseado em EEG

BARBOSA, André F. ; SOUZA, Bryan C. ; PEREIRA JUNIOR, Antônio ; MEDEIROS, Adelardo A. D.de, . Implementação de Classificador de Tarefas Mentais Baseado em EEG. In: CONGRESSO BRASILEIRO DE REDES NEURAIS, 9., 2009, Ouro Preto, MG. Anais... Ouro Preto, MG, 2009

Layering the Internet-of-Things with Multicasting in Flow-sensors for Internet-of-services

Development of Internet-of-Services will be hampered by heterogeneous Internet-of-Things infrastructures, such as inconsistency in communicating with participating objects, connectivity between them, topology definition & data transfer, access via cloud computing for data storage etc. Our proposed solutions are applicable to a random topology scenario that allow establishing of multi-operational sensor networks out of single networks and/or single service networks with the participation of multiple networks; thus allowing virtual links to be created and resources to be shared. The designed layers are context-aware, application-oriented, and capable of representing physical objects to a management system, along with discovery of services. The reliability issue is addressed by deploying IETF supported IEEE 802.15.4 network model for low-rate wireless personal networks. Flow- sensor succeeded better results in comparison to the typical - sensor from reachability, throughput, energy consumption and diversity gain viewpoint and through allowing the multicast groups into maximum number, performances can be improved.

A survey of intrusion detection system technologies

This paper provides an overview of IDS types and how they work as well as configuration considerations and issues that affect them. Advanced methods of increasing the performance of an IDS are explored such as specification based IDS for protecting Supervisory Control And Data Acquisition (SCADA) and Cloud networks. Also by providing a review of varied studies ranging from issues in configuration and specific problems to custom techniques and cutting edge studies a reference can be provided to others interested in learning about and developing IDS solutions. Intrusion Detection is an area of much required study to provide solutions to satisfy evolving services and networks and systems that support them. This paper aims to be a reference for IDS technologies other researchers and developers interested in the field of intrusion detection.

Implementação de classificador de tarefas mentais baseado em EEG

BARBOSA, André F. ; SOUZA, Bryan C. ; PEREIRA JUNIOR, Antônio ; MEDEIROS, Adelardo A. D.de, . Implementação de Classificador de Tarefas Mentais Baseado em EEG. In: CONGRESSO BRASILEIRO DE REDES NEURAIS, 9., 2009, Ouro Preto, MG. Anais... Ouro Preto, MG, 2009

Electrical and optical studies on carbon nanotube arrays

Single-walled carbon nanotubes (SWNTs) have been studied as a prominent class of high performance electronic materials for next generation electronics. Their geometry dependent electronic structure, ballistic transport and low power dissipation due to quasi one dimensional transport, and their capability of carrying high current densities are some of the main reasons for the optimistic expectations on SWNTs. However, device applications of individual SWNTs have been hindered by uncontrolled variations in characteristics and lack of scalable methods to integrate SWNTs into electronic devices. One relatively new direction in SWNT electronics, which avoids these issues, is using arrays of SWNTs, where the ensemble average may provide uniformity from device to device, and this new breed of electronic material can be integrated into electronic devices in a scalable fashion. This dissertation describes (1) methods for characterization of SWNT arrays, (2) how the electrical transport in these two-dimensional arrays depend on length scales and spatial anisotropy, (3) the interaction of aligned SWNTs with the underlying substrate, and (4) methods for scalable integration of SWNT arrays into electronic devices. The electrical characterization of SWNT arrays have been realized by polymer electrolyte-gated SWNT thin film transistors (TFTs). Polymer electrolyte-gating addresses many technical difficulties inherent to electrical characterization by gating through oxide-dielectrics. Having shown polymer electrolyte-gating can be successfully applied on SWNT arrays, we have studied the length scaling dependence of electrical transport in SWNT arrays. Ultrathin films formed by sub-monolayer surface coverage of SWNT arrays are very interesting systems in terms of the physics of two-dimensional electronic transport. We have observed that they behave qualitatively different than the classical conducting films, which obey the Ohm’s law. The resistance of an ultrathin film of SWNT arrays is indeed non-linear with the length of the film, across which the transport occurs. More interestingly, a transition between conducting and insulating states is observed at a critical surface coverage, which is called percolation limit. The surface coverage of conducting SWNTs can be manipulated by turning on and off the semiconductors in the SWNT array, leading to the operation principle of SWNT TFTs. The percolation limit depends also on the length and the spatial orientation of SWNTs. We have also observed that the percolation limit increases abruptly for aligned arrays of SWNTs, which are grown on single crystal quartz substrates. In this dissertation, we also compare our experimental results with a two-dimensional stick network model, which gives a good qualitative picture of the electrical transport in SWNT arrays in terms of surface coverage, length scaling, and spatial orientation, and briefly discuss the validity of this model. However, the electronic properties of SWNT arrays are not only determined by geometrical arguments. The contact resistances at the nanotube-nanotube and nanotube-electrode (bulk metal) interfaces, and interactions with the local chemical groups and the underlying substrates are among other issues related to the electronic transport in SWNT arrays. Different aspects of these factors have been studied in detail by many groups. In fact, I have also included a brief discussion about electron injection onto semiconducting SWNTs by polymer dopants. On the other hand, we have compared the substrate-SWNT interactions for isotropic (in two dimensions) arrays of SWNTs grown on Si/SiO2 substrates and horizontally (on substrate) aligned arrays of SWNTs grown on single crystal quartz substrates. The anisotropic interactions associated with the quartz lattice between quartz and SWNTs that allow near perfect horizontal alignment on substrate along a particular crystallographic direction is examined by Raman spectroscopy, and shown to lead to uniaxial compressive strain in as-grown SWNTs on single crystal quartz. This is the first experimental demonstration of the hard-to-achieve uniaxial compression of SWNTs. Temperature dependence of Raman G-band spectra along the length of individual nanotubes reveals that the compressive strain is non-uniform and can be larger than 1% locally at room temperature. Effects of device fabrication steps on the non-uniform strain are also examined and implications on electrical performance are discussed. Based on our findings, there are discussions about device performances and designs included in this dissertation. The channel length dependences of device mobilities and on/off ratios are included for SWNT TFTs. Time response of polymer-electrolyte gated SWNT TFTs has been measured to be ~300 Hz, and a proof-of-concept logic inverter has been fabricated by using polymer electrolyte gated SWNT TFTs for macroelectronic applications. Finally, I dedicated a chapter on scalable device designs based on aligned arrays of SWNTs, including a design for SWNT memory devices.

'Explicit Learning: an Effort towards Human Scheduling Algorithms'

Abstract Scheduling problems are generally NP-hard combinatorial problems, and a lot of research has been done to solve these problems heuristically. However, most of the previous approaches are problem-specific and research into the development of a general scheduling algorithm is still in its infancy. Mimicking the natural evolutionary process of the survival of the fittest, Genetic Algorithms (GAs) have attracted much attention in solving difficult scheduling problems in recent years. Some obstacles exist when using GAs: there is no canonical mechanism to deal with constraints, which are commonly met in most real-world scheduling problems, and small changes to a solution are difficult. To overcome both difficulties, indirect approaches have been presented (in [1] and [2]) for nurse scheduling and driver scheduling, where GAs are used by mapping the solution space, and separate decoding routines then build solutions to the original problem. In our previous indirect GAs, learning is implicit and is restricted to the efficient adjustment of weights for a set of rules that are used to construct schedules. The major limitation of those approaches is that they learn in a non-human way: like most existing construction algorithms, once the best weight combination is found, the rules used in the construction process are fixed at each iteration. However, normally a long sequence of moves is needed to construct a schedule and using fixed rules at each move is thus unreasonable and not coherent with human learning processes. When a human scheduler is working, he normally builds a schedule step by step 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 completed yet, thus having the ability to finish a schedule by using flexible, rather than fixed, rules. In this research we intend to design more human-like scheduling algorithms, by using ideas derived from Bayesian Optimization Algorithms (BOA) and Learning Classifier Systems (LCS) to implement explicit learning from past solutions. BOA can be applied to learn to identify good partial solutions and to complete them by building a Bayesian network of the joint distribution of solutions [3]. A Bayesian network is a directed acyclic graph with each node corresponding to one variable, and each variable corresponding to individual rule by which a schedule will be constructed step by step. The conditional probabilities are computed according to an initial set of promising solutions. Subsequently, each new instance for each node is generated by using the corresponding conditional probabilities, until values for all nodes have been generated. 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 Bayesian network is updated again using the current set of good rule strings. The algorithm thereby tries to explicitly identify and mix promising building blocks. It should be noted that for most scheduling problems the structure of the network model is known and all the 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. In the LCS approach, each rule has its strength showing its current usefulness in the system, and this strength is constantly assessed [4]. To implement sophisticated learning based on previous solutions, an improved LCS-based algorithm is designed, which consists of the following three steps. The initialization step is to assign each rule at each stage a constant initial strength. Then rules are selected by using the Roulette Wheel strategy. The next step is to reinforce the strengths of the rules used in the previous solution, keeping the strength of unused rules unchanged. The selection step is to select fitter rules for the next generation. It is envisaged that the LCS part of the algorithm will be used as a hill climber to the BOA algorithm. This is exciting and ambitious research, which might provide the stepping-stone for a new class of scheduling algorithms. Data sets from nurse scheduling and mall problems will be used as test-beds. It is envisaged that once the concept has been proven successful, it will be implemented into general scheduling algorithms. It is also hoped that this research will give some preliminary answers about how to include human-like learning into scheduling algorithms and may therefore be of interest to researchers and practitioners in areas of scheduling and evolutionary computation. References 1. Aickelin, U. and Dowsland, K. (2003) 'Indirect Genetic Algorithm for a Nurse Scheduling Problem', Computer & Operational Research (in print). 2. Li, J. and Kwan, R.S.K. (2003), 'Fuzzy Genetic Algorithm for Driver Scheduling', European Journal of Operational Research 147(2): 334-344. 3. Pelikan, M., Goldberg, D. and Cantu-Paz, E. (1999) 'BOA: The Bayesian Optimization Algorithm', IlliGAL Report No 99003, University of Illinois. 4. Wilson, S. (1994) 'ZCS: A Zeroth-level Classifier System', Evolutionary Computation 2(1), pp 1-18.

'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,

Superconducting Transition-Edge Sensor Physics

Despite record-setting performance demonstrated by superconducting Transition Edge Sensors (TESs) and growing utilization of the technology, a theoretical model of the physics governing TES devices superconducting phase transition has proven elusive. Earlier attempts to describe TESs assumed them to be uniform superconductors. Sadleir et al. 2010 shows that TESs are weak links and that the superconducting order parameter strength has significant spatial variation. Measurements are presented of the temperature T and magnetic field B dependence of the critical current Ic measured over 7 orders of magnitude on square Mo/Au bilayers ranging in length from 8 to 290 microns. We find our measurements have a natural explanation in terms of a spatially varying order parameter that is enhanced in proximity to the higher transition temperature superconducting leads (the longitudinal proximity effect) and suppressed in proximity to the added normal metal structures (the lateral inverse proximity effect). These in-plane proximity effects and scaling relations are observed over unprecedentedly long lengths (in excess of 1000 times the mean free path) and explained in terms of a Ginzburg-Landau model. Our low temperature Ic(B) measurements are found to agree with a general derivation of a superconducting strip with an edge or geometric barrier to vortex entry and we also derive two conditions that lead to Ic rectification. At high temperatures the Ic(B) exhibits distinct Josephson effect behavior over long length scales and following functional dependences not previously reported. We also investigate how film stress changes the transition, explain some transition features in terms of a nonequilibrium superconductivity effect, and show that our measurements of the resistive transition are not consistent with a percolating resistor network model.

Scheduling of track inspection and maintenance activities in railroad networks

The U.S. railroad companies spend billions of dollars every year on railroad track maintenance in order to ensure safety and operational efficiency of their railroad networks. Besides maintenance costs, other costs such as train accident costs, train and shipment delay costs and rolling stock maintenance costs are also closely related to track maintenance activities. Optimizing the track maintenance process on the extensive railroad networks is a very complex problem with major cost implications. Currently, the decision making process for track maintenance planning is largely manual and primarily relies on the knowledge and judgment of experts. There is considerable potential to improve the process by using operations research techniques to develop solutions to the optimization problems on track maintenance. In this dissertation study, we propose a range of mathematical models and solution algorithms for three network-level scheduling problems on track maintenance: track inspection scheduling problem (TISP), production team scheduling problem (PTSP) and job-to-project clustering problem (JTPCP). TISP involves a set of inspection teams which travel over the railroad network to identify track defects. It is a large-scale routing and scheduling problem where thousands of tasks are to be scheduled subject to many difficult side constraints such as periodicity constraints and discrete working time constraints. A vehicle routing problem formulation was proposed for TISP, and a customized heuristic algorithm was developed to solve the model. The algorithm iteratively applies a constructive heuristic and a local search algorithm in an incremental scheduling horizon framework. The proposed model and algorithm have been adopted by a Class I railroad in its decision making process. Real-world case studies show the proposed approach outperforms the manual approach in short-term scheduling and can be used to conduct long-term what-if analyses to yield managerial insights. PTSP schedules capital track maintenance projects, which are the largest track maintenance activities and account for the majority of railroad capital spending. A time-space network model was proposed to formulate PTSP. More than ten types of side constraints were considered in the model, including very complex constraints such as mutual exclusion constraints and consecution constraints. A multiple neighborhood search algorithm, including a decomposition and restriction search and a block-interchange search, was developed to solve the model. Various performance enhancement techniques, such as data reduction, augmented cost function and subproblem prioritization, were developed to improve the algorithm. The proposed approach has been adopted by a Class I railroad for two years. Our numerical results show the model solutions are able to satisfy all hard constraints and most soft constraints. Compared with the existing manual procedure, the proposed approach is able to bring significant cost savings and operational efficiency improvement. JTPCP is an intermediate problem between TISP and PTSP. It focuses on clustering thousands of capital track maintenance jobs (based on the defects identified in track inspection) into projects so that the projects can be scheduled in PTSP. A vehicle routing problem based model and a multiple-step heuristic algorithm were developed to solve this problem. Various side constraints such as mutual exclusion constraints and rounding constraints were considered. The proposed approach has been applied in practice and has shown good performance in both solution quality and efficiency.

'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,