There have recently been significant leaps in deep reinforcement learning algorithms, with notable successes in games such as Atari arcade games and Go; however, there is still a need to adapt these techniques to be more widely applicable in other domains, such as the life science sector. Identifying regulatory relationships between genes is one of the primary research activities carried out by molecular biologists and geneticists, since learning the structure of gene regulatory networks is critical for many applications, for example understanding the origins of many diseases and how crops respond to their environments. Biologists sequentially conduct experiments that provide information about the gene network structure, but they must operate under strict cost and time limits. This project aims to formulate this experiment design procedure in a reinforcement-learning framework, to ascertain how biologists should prioritise experiments to maximise information about the gene networks, under constraints. The primary deliverable will be a Computer-aided Experimental Design (CoED) software tool to aid researchers in utilising their resources most effectively. This reinforcement-learning framework could also be used to identify the bottlenecks for biomedical research, such as the pricing model or the time-intensity of certain experiments, thereby identifying the most impactful areas for further development in experimental methodology. We will deliver impact by providing consultation services to laboratory supply and service providers, and through our collaboration with our industrial partner Google Brain Genomics. This project primarily aligns with the new approaches to data science and high productivity services through specialised artificial intelligence priority areas of this call.

One of the distinguishing characteristics of software systems is that they evolve: new patches are committed to software repositories and new versions are released to users on a continuous basis. Unfortunately, many of these changes bring unexpected failures that break the stability of the system or affect its security, and users face the uncomfortable choice between using an old stable version which misses recent features and bug fixes, and upgrading to a new version which improves the software in certain ways, only to introduce other bugs and security vulnerabilities. In this fellowship, I plan to investigate novel techniques for improving the reliability and security of evolving software, based on the idea of combining the execution of multiple software versions in such a way as to increase the reliability and security of the "multi-version" application and eliminate a large number of common bugs introduced by software updates. This is an ambitious proposal, which presents several challenges spanning the areas of software engineering, computer systems, and security: understanding how software evolves, and particularly the effects of incorrect updates on software evolution; addressing the technical challenges of multi-version execution such as creating an application-level sandboxing environment and devising lightweight record and replay techniques; designing error recovery strategies that effectively combine different software versions; and determining the applicability of multi-version execution to the different types of applications and code changes encountered in practice.

This project is a direct response to significant changes taking place in the domain of computing and the arts. Recent developments in Artificial Intelligence and Machine Learning are leading to a revolution in how music and art is being created by researchers (Broad and Grierson, 2016). However, this technology has not yet been integrated into software aimed at creatives. Due to the complexities of machine learning, and the lack of usable tools, such approaches are only usable by experts. In order to address this, we will create new, user-friendly technologies that enable the lay user - composers as well as amateur musicians - to understand and apply these new computational techniques in their own creative work. The potential for machine learning to support creative activity is increasing at a significant rate, both in terms of creative understanding and potential applications. Emerging work in the field of music and sound generation extends from musical robots to generative apps, and from advanced machine listening to devices that can compose in any given style. By leveraging the internet as a live software ecosystem, the proposed project examines how such technology can best reach artists, and live up to its potential to fundamentally change creative practice in the field. Rather than focussing on the computer as an original creator, we will create platforms where the newest techniques can be used by artists as part of their day-to-day creative practices. Current research in artificial intelligence, and in particular machine learning, have led to an incredible leap forward in the performance of AI systems in areas such as speech and image recognition (Cortana, Siri etc.). Google and others have demonstrated how these approaches can be used for creative purposes, including the generation of speech and music (DeepMinds's WaveNet and Google's Magenta), images (Deep Dream) and game intelligence (DeepMind's AlphaGo). The investigators in this project have been using Deep Learning, Convolutional Neural Networks (CNNs), Recurrent Neural Networks (RNNs), Long Short-Term Memory Networks (LSTMs), and other approaches to develop intelligent systems that can be used by artists to create sound and music. We are already among the first in the world to create reusable software that can 'listen' to large amounts of sound recordings, and use these as examples to create entirely new recordings at the level of audio. Our systems produce outcomes that out-perform many other previously funded research outputs in these areas. In this three-year project, we will develop and disseminate creative systems that can be used by musicians and artists in the creation of entirely new music and sound. We will show how such approaches can affect the future of other forms of media, such as film and the visual arts. We will do so by developing a creative platform, using the most accessible public forum available: the World Wide Web. We will achieve this through development of a high level live coding language for novice users, with simplified metaphors for the understanding of complex techniques including deep learning. We will also release the machine learning libraries we create for more advanced users who want to use machine learning technology as part of their creative tools. The project will involve end-users throughout, incorporating graduate students, professional artists, and participants in online learning environments. We will disseminate our work early, gaining the essential feedback required to deliver a solid final product and outcome. The efficacy of such techniques has been demonstrated with systems such as Sonic Pi and Ixi Lang, within a research domain already supported by the AHRC through the Live Coding Network (AH/L007266/1), and by EC in the H2020 project, RAPID-MIX. Finally, this research will strongly contribute to dialogues surrounding the future of music and the arts, consolidating the UK's leadership in these fields.

The repeated occurrence of high profile flood events, both within the UK (e.g. Somerset Levels, 2014) and internationally (e.g. Australia and Thailand, 2011), has resulted in sustained public, commercial, political and scientific interest in flood risk. The heightened profile of flooding within the scientific community is reflected in the number of recent and current NERC thematic programmes dedicated wholly or in large part to this topic. The Hydrology Group at the University of Bristol has been involved in or led numerous programmes, with the result being a substantial improvement in our understanding of flood risk in data-rich regions such as the UK. However flooding is very clearly a worldwide phenomenon and consistent global hazard and risk products do not currently exist but are very urgently required by international governments and NGOs, and by British businesses. Looking to the future, the continued expansion of cities located on river floodplains and coastal deltas due to population growth and migration is expected to produce a significant increase in flood risk over the coming decades. The prospect is of particular concern to the (re)insurance market, a critical component of the UK financial services sector, as well as to global humanitarian and development organisations such as the World Bank and the United Nations World Food Programme. Recent losses have proven significant: household losses resulting from the summer 2007 floods in the UK reached £2.5 billion, with business losses accounting for a further £1 billion. The reinsurance firm Munich Re estimates that total economic losses from the Australian and Thailand events of 2011 were USD 2.8 billion and USD 40 billion respectively. Against this background and despite being largely focussed on UK flood risk the above mentioned recent NERC-funded programmes have yielded a number of important research developments that have started to close the knowledge gap that exists for global flood risk. Perhaps most critically, NERC funded research undertaken at the University of Bristol has yielded algorithms that allow very large scale flood models to be undertaken. Further NERC funded research has guided the development of model structures that allow models with global coverage to be developed. The Hydrology Group at the University of Bristol has built strong links with the insurance market through a long standing partnership with the Willis Research Network, Willis being one of the three dominant global reinsurance brokers. Through this relationship with Willis and other industry partners the potential to develop both freely available global flood hazard layers and commercial flood risk products was identified. Currently flood risk products appropriate for insurance and re-insurance markets exist in only a few territories worldwide. For many developed and emerging markets is near zero provision but significant and growing exposure. At the same time, freely available global flood hazard layers would be a significant benefit to humanitarian agencies such as the World Food Programme and international organisations with significant interests in risk reduction such as the World Bank. Therefore, this project will produce global flood hazard layers using state of the art methods developed as part of NERC funded science and provide massively open access to these via Google Earth. These layers will be the first global openly accessible assessment of flood hazard that can be used by Governments, NGOs and the entire global public thereby changing fundamentally global awareness of flood hazard and leading to a step change in flood resilience. We will then work with the global (re)insurance industry to identify commercial opportunities relating to this and related NERC funded science. The proposal is supported by Google, NASA-JPL, World Bank, World Food Programme, Willis, Mitsui Sumitomo, Montpelier-Re and the Universities of Reading and Newcastle.

In digital discrimination, users are treated unfairly, unethically or just differently based on their personal data. Examples include low-income neighborhoods targeted with high-interest loans; women being undervalued by 21% in online marketing; and online ads suggestive of arrest records appearing more often with searches of black-sounding names than white-sounding names. Digital discrimination very often reproduces existing instances of discrimination in the offline world by either inheriting the biases of prior decision makers, or simply reflecting widespread prejudices in society. Digital discrimination may also have an even more perverse result, it may exacerbate existing inequalities by causing less favourable treatment for historically disadvantaged groups, suggesting they actually deserve that treatment. As more and more tasks are delegated to computers, mobile devices, and autonomous systems, digital discrimination is becoming a huge problem. Digital discrimination can be the result of algorithmic biases, i.e., the way in which a particular algorithm has been designed creates discriminatory outcomes, but it also occurs using non-biased algorithms when they are fed or trained with biased data. Research has been conducted on so-called fair algorithms, tackling biased input data, demonstrating learned biases, and measuring relative influence of data attributes, which can quantify and limit the extent of bias introduced by an algorithm or dataset. But, how much bias is too much? That is, what is legal, ethical and/or socially-acceptable? And even more importantly, how do we translate those legal, ethical, or social expectations into automated methods that attest digital discrimination in datasets and algorithms? DADD (Discovering and Attesting Digital Discrimination) is a *novel cross-disciplinary collaboration* to address these open research questions following a continuously-running co-creation process with academic (Computer Science, Digital Humanities, Law and Ethics) and non-academic partners (Google, AI Club), and the general public, including technical and non-technical users. DADD will design ground-breaking methods to certify whether or not datasets and algorithms discriminate by automatically verifying computational non-discrimination norms, which will in turn be formalised based on socio-economic, cultural, legal, and ethical dimensions, creating the new *transdisciplinary field of digital discrimination certification*.