The activity of a biological cell is regulated by interactions between genes and proteins. In artificial intelligence, this has led to the creation of developmental gene regulatory network (GRN) models which aim to exploit these mechanisms to algorithmically build complex designs. The emerging field of GRNs for control aims to instead exploit these natural mechanisms and this ability to encode a large variety of behaviours within a single evolvable genetic program for the solution of control problems. This work aims to extend the application domain of GRN models to previously unsolved control problems; the focus will here be on reinforcement learning problems, in which the dynamics of the system controlled are kept from the controller and only sparse feedback is given to it. This category of problems closely matches the challenges faced by natural evolution in generating biological GRNs. Starting with an existing GRN model, the fractal GRN (FGRN) model, a successful application to a standard control problem will be presented, followed by multiple improvements to the FGRN model and its associated genetic algorithm, resulting in better performances in terms of both reliability and speed. Limitations will be identified in the FGRN model, leading to the introduction of the Input-Merge- Regulate-Output (IMRO) architecture for GRN models, an implementation of which will show both quantitative and qualitative improvements over the FGRN model, solving harder control problems. The resulting model also displays useful features which should facilitate further extension and real-world use of the system.

Experimental characterisation of protein cellular function can be prohibitively expensive and take years to complete. To address this problem, this thesis focuses on the development of computational approaches to predict function from sequence. For sequences with well characterised close relatives, annotation is trivial, orphans or distant homologues present a greater challenge. The use of a feature based method employing ensemble support vector machines to predict individual Gene Ontology classes is investigated. It is found that different combinations of feature inputs are required to recognise different functions. Although the approach is applicable to any human protein sequence, it is restricted to broadly descriptive functions. The method is well suited to prioritisation of candidate functions for novel proteins rather than to make highly accurate class assignments. Signatures of common function can be derived from different biological characteristics; interactions and binding events as well as expression behaviour. To investigate the hypothesis that common function can be derived from expression information, public domain human microarray datasets are assembled. The questions of how best to integrate these datasets and derive features that are useful in function prediction are addressed. Both co-expression and abundance information is represented between and within experiments and investigated for correlation with function. It is found that features derived from expression data serve as a weak but significant signal for recognising functions. This signal is stronger for biological processes than molecular function categories and independent of homology information. The protein domain has historically been coined as a modular evolutionary unit of protein function. The occurrence of domains that can be linked by ancestral fusion events serves as a signal for domain-domain interactions. To exploit this information for function prediction, novel domain architecture and fused architecture scores are developed. Architecture scores rather than single domain scores correlate more strongly with function, and both architecture and fusion scores correlate more strongly with molecular functions than biological processes. The final study details the development of a novel heterogeneous function prediction approach designed to target the annotation of both homologous and non-homologous proteins. Support vector regression is used to combine pair-wise sequence features with expression scores and domain architecture scores to rank protein pairs in terms of their functional similarities. The target of the regression models represents the continuum of protein function space empirically derived from the Gene Ontology molecular function and biological process graphs. The merit and performance of the approach is demonstrated using homologous and non-homologous test datasets and significantly improves upon classical nearest neighbour annotation transfer by sequence methods. The final model represents a method that achieves a compromise between high specificity and sensitivity for all human proteins regardless of their homology status. It is expected that this strategy will allow for more comprehensive and accurate annotations of the human proteome.

In recent years, some trends have emerged that pertain both to mobile devices and the Web. On one side, mobile devices have transitioned from being simple wireless phones to become ubiquitous Web-enabled users' companions. On the other side, the Web has evolved from an online one-size-fits-all collection of interlinked documents to become an open platform of personalised services and content. It will not be long before these trends will converge and create a Seamless Web: an integrated environment where, besides traditional services delivered by powerful server machines accessible via wide area networks, new services and content will be offered by users to users via their portable devices. As a result, mobile users will soon be exposed - in addition to traditional "on-line" Web services/content - to a parallel universe of pervasive "off-line" services provided by devices in their surroundings. Such circumstances will raise new challenges when it comes to selecting the services to rely on, that will require solutions grounded on the characteristics of mobile environments. Two aspects will require particular attention: first, users will have access to a countless multitude of services impossible to explore; they will need assistance to identify, among this multitude, those services they are most likely to enjoy. Secondly, if today's services (and their providers) are always-on, `static' and aiming at Five 9s availability, tomorrow's pervasive services will be mobile (as devices move), fine-grained, increasingly composite (to provide richer functionalities) and so more unreliable by nature. Our research tackles the problem of service discovery in pervasive environments in two ways: on one hand, we support personalised discovery by means of a mobile recommender system, easing the discovery of pervasive services appealing to end-users. On the other hand, we enable reliable discovery, by reasoning on the composite nature of pervasive services and the physical availability of their component providers. Overall, we provide a discovery method that enables 'better' pervasive services, where by 'better' we mean both `more interesting' to the user and 'more reliable'.

The usage of machine learning techniques for the prediction of financial time series is investigated. Both discriminative and generative methods are considered and compared to more standard financial prediction techniques. Generative methods such as Switching Autoregressive Hidden Markov and changepoint models are found to be unsuccessful at predicting daily and minutely prices from a wide range of asset classes. Committees of discriminative techniques (Support Vector Machines (SVM), Relevance Vector Machines and Neural Networks) are found to perform well when incorporating sophisticated exogenous financial information in order to predict daily FX carry basket returns. The higher dimensionality that Electronic Communication Networks make available through order book data is transformed into simple features. These volume-based features, along with other price-based ones motivated by common trading rules, are used by Multiple Kernel Learning (MKL) to classify the direction of price movement for a currency over a range of time horizons. Outperformance relative to both individual SVM and benchmarks is found, along with an indication of which features are the most informative for financial prediction tasks. Fisher kernels based on three popular market microstructural models are added to the MKL set. Two subsets of this full set, constructed from the most frequently selected and highest performing individual kernels are also investigated. Furthermore, kernel learning is employed - optimising hyperparameter and Fisher feature parameters with the aim of improving predictive performance. Significant improvements in out-of-sample predictive accuracy relative to both individual SVM and standard MKL is found using these various novel enhancements to the MKL algorithm.

Networks are important. They provide a general framework for representing the relationships or interactions which impose dependencies between entities. Network nodes represent entities and the links between them represent relations/interactions. Nodes' attributes contain non-structural features of the entities. For example, in a social network where nodes represent people and links represent friendships, attributes could represent features such as age, race or gender. In many situations it is easy to observe the link structure of a network, but not the attributes of the nodes. For example, in an on-line social network it may be possible to observe all friendships, but the observability of attributes are determined by the user's privacy settings. As a result the state-of-the-art of learning in networks tend to either focus on clustering nodes with similar link patterns (i.e functional communities) or predicting the missing attribute of nodes (i.e. node labels). In this work we bring these two ideas together to examine how the identification of communities and related structures can be used to predict the hidden attributes of the network nodes. The models we present are effective, flexible and principled. They are effective in their ability to predict discrete (either binary or multivalued) and continuous node attributes in real world network datasets. They are flexible in that they can adapt to and identify a wide range of network structures. They are principled in that they are based on sound theoretical methods of Bayesian statistics. We achieve this by a series of novel extensions to the stochastic blockmodel, a probabilistic generative model for identifying functional communities.

This thesis offers some new analyses and presents some new methods for learning in the context of exploiting structure defined by data – for example, when a data distribution has a submanifold support, exhibits cluster structure or exists as an object such as a graph. 1. We present a new PAC-Bayes analysis of learning in this context, which is sharp and in some ways presents a better solution than uniform convergence methods. The PAC-Bayes prior over a hypothesis class is defined in terms of the unknown true risk and smoothness of hypotheses w.r.t. the unknown data-generating distribution. The analysis is “localized” in the sense that complexity of the model enters not as the complexity of an entire hypothesis class, but focused on functions of ultimate interest. Such bounds are derived for various algorithms including SVMs. 2. We consider an idea similar to the p-norm Perceptron for building classifiers on graphs. We define p-norms on the space of functions over graph vertices and consider interpolation using the pnorm as a smoothness measure. The method exploits cluster structure and attains a mistake bound logarithmic in the diameter, compared to a linear lower bound for standard methods. 3. Rademacher complexity is related to cluster structure in data, quantifying the notion that when data clusters we can learn well with fewer examples. In particular we relate transductive learning to cluster structure in the empirical resistance metric. 4. Typical methods for learning over a graph do not scale well in the number of data points – often a graph Laplacian must be inverted which becomes computationally intractable for large data sets. We present online algorithms which, by simplifying the graph in principled way, are able to exploit the structure while remaining computationally tractable for large datasets. We prove state-of-the-art performance guarantees.

The peer-to-peer paradigm is an important alternative to the traditional client-server model in computer networks, making up a significant share of the bandwidth used globally. In client-server scenarios there usually is an external reason of why the server provides its service to the clients. But there usually is no external incentive in peer-to-peer networks to share data. In fact there are two good reasons not to. Firstly, providing data to another node consumes bandwidth, which will always be limited and whose use may might incur a cost. Secondly, the process of making data accessible is also costly. The data needs to be obtained, its existence needs to be advertised and individuals need to decide which data to share. An incentive is required for nodes to offer their resources. We propose a generalisation of the BitTorrent incentives mechanism that improves it in two important ways. It works for a broader range of files in terms of size and popularity, enabling a simple BitTorrent-like tit-for-tat incentives mechanism for files that do not work with BitTorrent. At the same time it provides peers with an incentive to share more files. In BitTorrent, peers download pieces of the same file from each other. This is a bartering ring of length 2. Our algorithm extends this idea by allowing pieces of different files to be exchanged and by allowing longer rings with more nodes to be formed. For this, rings need to be identified in an overlay graph that consists of the nodes and potential downloads among them. But no node has knowledge of the graph other than its direct neighbours. For the incentives mechanism to work once rings have been identified, a group consensus needs to be reached to start the downloads. We propose distributed algorithms for these problems and evaluate them experimentally using a simulation. We are able to show that in some cases our incentives algorithm works better for small and rare files than BitTorrent.

Being able to select decoy structures that are closest to the native one is essential to any folding simulation. Indeed, modern algorithms use heuristics to quickly sample the conformational space, and as such, will generate a large number of candidate structures. In this thesis, we create a new statistical energy function to correctly discriminate near-native decoy structures, using three complementary approaches to derive energies from known conformations and decoys. First, we used a classical definition, where the observed state is modelled by taking a set of 1078 short, well-resolved, non-redundant crystal structures from the PDB, and the reference state is taken as the distribution expected at random. In our second method, which we call “hybrid”, we used the native structures as the observed state, just as in the classical formulation, but this time using the worse generated decoys as the reference state. Finally, our third method, called “decoy-based”, uses only decoys, taking the better than average models as the observed state, and the worse than average as the reference state. Using the three methods above, we generated potentials to model solvation, hydrogen bonding, and pairwise atomic distances and orientation. We found that overall, combining solvation, atomic distance and orientation using the decoy-based method produced the best results, with a 10% enrichment score of 0.73 versus 0.51 for the classical formulation, and 0.41 for our benchmark potential, DFIRE2. Our final potential, called the DOS potential, was created by combining the classical, hybrid and decoy-based potentials, and achieved a 10% enrichment score of 0.75 versus 0.41 for DFIRE2.

Optimal control is a profound and fascinating subject that regularly attracts interest from numerous scien- tific disciplines, including both pure and applied Mathematics, Computer Science, Artificial Intelligence, Psychology, Neuroscience and Economics. In 1960 Rudolf Kalman discovered that there exists a dual- ity between the problems of filtering and optimal control in linear systems [84]. This is now regarded as a seminal piece of work and it has since motivated a large amount of research into the discovery of similar dualities between optimal control and statistical inference. This is especially true of recent years where there has been much research into recasting problems of optimal control into problems of statis- tical/approximate inference. Broadly speaking this is the perspective that we take in this work and in particular we present various applications of methods from the fields of statistical/approximate inference to optimal control, planning and Reinforcement Learning. Some of the methods would be more accu- rately described to originate from other fields of research, such as the dual decomposition techniques used in chapter(5) which originate from convex optimisation. However, the original motivation for the application of these techniques was from the field of approximate inference. The study of dualities be- tween optimal control and statistical inference has been a subject of research for over 50 years and we do not claim to encompass the entire subject. Instead, we present what we consider to be a range of interesting and novel applications from this field of research

Denial-of-Service attacks continue to grow in size and frequency despite serious underreporting. While several research solutions have been proposed over the years, they have had important deployment hurdles that have prevented them from seeing any significant level of deployment on the Internet. Commercial solutions exist, but they are costly and generally are not meant to scale to Internet-wide levels. In this thesis we present three filtering architectures against large Denial-of-Service attacks. Their emphasis is in providing an effective solution against such attacks while using simple mechanisms in order to overcome the deployment hurdles faced by other solutions. While these are well-suited to being implemented in fast routing hardware, in the early stages of deployment this is unlikely to be the case. Because of this, we implemented them on low-cost off-the-shelf hardware and evaluated their performance on a network testbed. The results are very encouraging: this setup allows us to forward traffic on a single PC at rates of millions of packets per second even for minimum-sized packets, while at the same time processing as many as one million filters; this gives us confidence that the architecture as a whole could combat even the large botnets currently being reported. Better yet, we show that this single-PC performance scales well with the number of CPU cores and network interfaces, which is promising for our solutions if we consider the current trend in processor design. In addition to using simple mechanisms, we discuss how the architectures provide clear incentives for ISPs that adopt them early, both at the destination as well as at the sources of attacks. The hope is that these will be sufficient to achieve some level of initial deployment. The larger goal is to have an architectural solution against large DoS deployed in place before even more harmful attacks take place; this thesis is hopefully a step in that direction.