Open Research Exeter - University of Exeter
Not a member yet
41213 research outputs found
Sort by
Real-Time World Crafting: Generating Structured Game Behaviors from Natural Language with Large Language Models
No full textWe present a novel architecture for safely integrating Large Language Models (LLMs) into interactive game engines, allowing players to "program" new behaviors using natural language. Our framework mitigates risks by using an LLM to translate commands into a constrained Domain-Specific Language (DSL), which configures a custom Entity-Component-System (ECS) at runtime. We evaluated this system in a 2D spell-crafting game prototype by experimentally assessing models from the Gemini, GPT, and Claude families with various prompting strategies. A validated LLM judge qualitatively rated the outputs, showing that while larger models better captured creative intent, the optimal prompting strategy is task-dependent: Chain-of-Thought improved creative alignment, while few-shot examples were necessary to generate more complex DSL scripts. This work offers a validated LLM-ECS pattern for emergent gameplay and a quantitative performance comparison for developers.</p
Scales of risk and adaptive ‘dread’: an evolutionary theory of risk inflation
No full textPeople often react to low probability, high damage events in which many die with strong avoidance behaviour. Indeed, analyses of behaviour following the September 11 terror attacks on New York City suggest that this caused a substantial number of additional, ‘indirect’ deaths as many people avoided flying for 12 months afterwards and took to the relatively risky highways of the US instead. We argue that such responses may have arisen as an adaptation to risks that wipe out a significant proportion of all carriers of an allele if they strike, e.g. storms. These are environmental fluctuations known as environmental or aggregate risks. At the opposite extreme, demographic risks affect individuals independently. We show that evolution by natural selection in fluctuating environments means it is adaptive to inflate environmental (aggregate) risks relative to demographic risks, where the inflation factor depends on the proportion of carriers of the allele that die if the risk strikes.</p
Ocean Meridional Overturning Circulation During the Early and Middle Miocene
No full textThe Miocene (∼23–5 Ma) is a past warm epoch when global surface temperatures varied between ∼5 and 8°C warmer than today, and CO2 concentration was ∼400–800 ppm. The narrowing/closing of the tropical ocean gateways and widening of high-latitude gateways throughout the Miocene is likely responsible for the evolution of the ocean's overturning circulation to its modern structure, though the mechanisms remain unclear. Here, we investigate early and middle Miocene ocean circulation in an opportunistic climate model intercomparison (MioMIP1), using 14 simulations with different paleogeography, CO2, and vegetation. The strength of the Southern Ocean-driven Meridional Overturning Circulation (SOMOC) bottom cell is similar in the Miocene and Pre-Industrial (PI) but dominates the Miocene global MOC due to weaker Northern Hemisphere overturning. The Miocene Atlantic MOC (AMOC) is weaker than PI in all the simulations (by 2–21 Sv), possibly due to its connection with an Arctic that is considerably fresher than today. Deep overturning in the North Pacific (PMOC) is present in three simulations (∼5–10 Sv), of which two have a weaker AMOC, and one has a stronger AMOC (compared to its PMOC). Surface freshwater fluxes control northern overturning such that the basin with the least freshwater gain has stronger overturning. While the orography, which impacts runoff direction (Pacific vs. Atlantic), has an inconsistent impact on northern overturning across simulations, overall, features associated with the early Miocene—such as a lower Tibetan Plateau, the Rocky Mountains, and a deeper Panama Seaway—seem to favor PMOC over AMOC.</p
Ancient dog mitogenomes support the dual dispersal of dogs and agriculture into South America
No full textArchaeological and palaeogenomic data show that dogs were the only domestic animals introduced during the early peopling of the Americas. Hunter-gatherer groups spread quickly towards the south of the continent, but it is unclear when dogs reached Central and South America. To address this issue, we generated and analysed 70 complete mitochondrial genomes from archaeological and modern dogs ranging from Central Mexico to Central Chile and Argentina, revealing the dynamics of dog populations. Our results demonstrate that pre-contact Central and South American dogs are all assigned to a specific clade that diverged after dogs entered North America. Specifically, the divergence time between North, Central and South American dog clades is consistent with the spread of agriculture and the adoption of maize in South America between 7000 and 5000 years ago. An isolation-by-distance best characterizes how dogs expanded into South America. We identify the arrival of new lineages of dogs in post-contact South America, likely of European origin, and their legacy in modern village dogs. Interestingly, the pre-contact Mesoamerican maternal origin of the Chihuahua has persisted in some modern individuals.</p
Active Variable Selection and Dimension Reduction Aided Design in History Matching
No full textIn this thesis, uncertainty quantification techniques such as Gaussian process emulation and history matching were applied to two complex simulators.
The first application, a high-dimensional output model which featured a collaboration between myself and other academics at Exeter UQ as part of a competition, resulting in a joint-second finish. This involved calibrating a multi-output model to synthetic
data using dimension reduction in output space. My attempt at this resulted in ruling out over 99% of input space before selecting a ‘best’ input value.
The second application, involved using a stochastic epidemiological model as part of the response to the outbreak of COVID-19. History matching was used to match the model to death data. The reductions in NROY space were negligible past a certain point. The results from this application form the two research chapters of my thesis.
The first research chapter conducts experiments with multiple variable selection methods within a history matching context. The effectiveness and consistency of these methods’ selections are tested and compared to the ‘true’ variables that should be selected using toy simulators. This is conducted over multiple history matching waves and design sizes. The results are largely inconclusive, in that it was difficult to determine across different toy simulators any trends relating to the correct variable selections being made.
The second research chapter concerns obtaining designs within areas of high emulator uncertainty within non-implausible regions of input space dictated by history matching. Dimension reduction is conducted on this non-implausible space using
weighted singular value decomposition prior to sampling any design. Whilst using dimension reduction did aid in ruling out more space, it performed poorly regarding emulator validation and reducing emulator uncertainty.</p
Rethinking Urban Water Network Design: A Reinforcement Learning Framework for Long-Term Flexible Planning
No full textEffectively planning the design of water distribution networks (WDNs) for their whole lifecycle is a complex task for water utilities due to the dynamic nature of WDNs, their long planning horizons, and the deep uncertainty that characterises key design parameters such as future water demand and population growth. Existing flexible design methods, which attempt to address these challenges, rely on static heuristic approaches and predefined decision pathways, requiring re-optimisation whenever new information becomes available. As additional scenarios are introduced, these methods also suffer from exponential increase in complexity, limiting their ability to adapt to emerging information and efficiently explore a wide range of future possibilities. In this work, we introduce a deep reinforcement learning (DRL) framework for the flexible, long-term design of WDNs. By formulating the least-cost staged design problem as a Markov Decision Process and training an agent using Proximal Policy Optimisation, our approach learns cost-effective, sequential interventions across multiple construction stages and future scenarios without relying on predefined decision trees. We evaluate our method on a modified New York Tunnels benchmark across three design tasks, ranging from static single-stage to flexible multi-stage design. Our results show that the DRL agent performs comparably to state-of-the-art heuristics for static and staged deterministic tasks. In the flexible design task, it autonomously devised adaptive strategies, clustered similar scenarios, and maintained high sample efficiency as the number of stages and scenarios increased. These findings highlight DRL as a promising alternative for the lifecycle design of WDNs, establishing a new paradigm for long-term water network planning under deep uncertainty.</p
Exploring the Role of Human Data in Data-Driven Design
No full textHuman data has significant value in Data-Driven Design, offering opportunities for user-centered product and service development. This paper explores how human data, categorized into behavioral, physiological, feedback, and emotional types, contributes to problem framing, iterative refinement, customization, and emotional design. Real-world case studies from academic literature and industry demonstrate how human data enables adaptive, personalized, and emotionally engaging solutions. Ethical challenges, including privacy, bias, and transparency, are explored, highlighting the importance of responsible data practices. The analysis underscores human data’s value in combining technical precision with empathetic design, fostering innovation and enhancing user experiences while promoting ethical use through principles of privacy, consent, and inclusivity.</p
Essays on employee welfare: Do firm-level risks and executive-level taxes matter for wage-related mistreatment, employee safety or pensions?
No full textThis PhD thesis comprises three empirical studies addressing a pivotal dimension of employee welfare: wage-related misconduct, safety and pensions. The first study, titled “Working amid climate uncertainties: The disciplining effect of climate risk on employee treatment”, examines the nexus between firm-level climate risk and wage theft, a major form of employee mistreatment. Examining this question is particularly important given the prevalence of wage theft, coupled with the inconclusive anecdotal reports and lack of empirical evidence on how climate risk may impact the phenomenon. Using extensive data on wage and hour violations, we provide robust evidence that wage theft is less prevalent in firms with high climate risk perception. To mitigate endogeneity concerns, we employ matched sample analysis, a test of coefficient stability, instrumental variable estimation, and difference-in-differences analysis around the adoption of the Paris Agreement. Additional analyses indicate that the negative relation between climate risk and wage theft is more pronounced in labour-intensive firms and firms that rely heavily on skilled labour, suggesting the role of human capital as the main channel. Further cross-sectional analyses reveal the results to be stronger for firms with high media coverage but attenuated in the presence of financial constraints and resilience against climate change. Finally, we demonstrate that heightened perceptions of climate risk are associated with the implementation of more robust labour management policies, while wage theft adversely impacts employee performance and productivity. Taken together, the findings of the study suggest that perceptions about climate risk can incentivise firms to exhibit improved employee treatment by cutting down on wage theft. Thus, it emphasises the significance of climate risk in shaping firms’ labour treatment policies.
The second empirical study, titled “CEO tax burden and workplace safety”, examines whether and how a CEO’s capital gains tax burden (that is, the tax liability arising on a CEO’s investment in the firm) impacts employees’ safety at the workplace. This is an important research question to examine in the light of the increasing social cost of workplace safety, and the recent findings in the literature highlighting the significant role CEO tax burden in corporate outcomes. Using comprehensive data on work-related injuries from the Occupational Safety and Health Administration, we find more cases of work-related injuries in establishments of firms led by CEOs with high capital gains tax burdens. The findings remain robust to several sensitivity analyses and tests designed to mitigate endogeneity concerns, including matched sample tests, instrumental variable analysis and a difference-in-differences analysis using the staggered reductions in state-level capital gains taxes. Further analyses suggest reductions in health and safety investments and increases in employee workloads as the potential mechanisms underscoring the results. In cross-sectional tests, we find that the observed relationship strengthens when CEOs have low levels of outside wealth and when firms are financially constrained. Moreover, we find that CEOs with high tax burdens compromise workplace safety, irrespective of whether these tax burdens are measured at the federal and state levels. However, the effect is more pronounced with state tax burdens. Finally, we provide evidence that CEO stock sales increase following a higher incidence of work-related injuries, suggesting that CEOs with high tax burdens may be compromising safety to unwound their locked-up stocks. Collectively, this study highlights the significant role of capital gains taxes in shaping employee welfare by demonstrating that the expected capital gains tax liabilities on CEOs’ investment in a firm can undermine employees’ safety at the workplace.
The final empirical chapter, titled “Firm-level political risk and pension buyout”, investigates whether the political risk faced by sponsor firms affects the likelihood of pension buyout transactions. Pension buyouts allow defined benefit plan sponsors to shift pension obligations to third-party insurers through a bulk annuity contract in exchange for a premium. Given the increasing concerns about political risk among DB plan sponsors, it is interesting to examine how this risk may affect the likelihood of pension buyouts. Using a sample of U.S DB plan sponsors from 2012 to 2021, we find a lower likelihood of pension buyout transactions when firms are exposed to a high political risk. The findings support the notion that corporations reduce immediate excessive spending in response to political risk. To address endogeneity concerns, we subject the findings to (1) a matched sample analysis, (2) a test involving only firms with a buyout record, (3) a falsification test, (4) a test of coefficient stability and (5) an instrumental variable analysis. The documented results persist through these tests, confirming the reliability of our findings. Cross-sectional tests indicate that the reducing effect of political risk on the likelihood of pension buyout is more pronounced in the presence of financing constraints, union coverage, and during the Obama administration. Evidence from supplemental tests also suggests that political risk negatively affects overall derisking and the extent to which employees and pension assets may be covered in a buyout transaction. Overall, the findings of this study underscore the significance of political risk in shaping corporate decisions regarding employee pensions.
In summary, this PhD thesis highlights how climate-related risk, managerial capital gains tax burden and political risk shape different dimensions of employee welfare.</p
Review article: AntArchitecture – building an age–depth model from Antarctica's radiostratigraphy to explore ice-sheet evolution
No full text Radio-echo sounding (RES) has revealed an internal architecture within both the West and East Antarctic ice sheets that records their depositional, deformational and melting histories. Crucially, RES-imaged internal-reflecting horizons, tied to ice-core age–depth profiles, can be treated as isochrones that record the age–depth structure across the Antarctic ice sheets. These enable the reconstruction of past climate and ice dynamical processes on large scales, which are complementary to but more spatially extensive than commonly used proxy records (e.g. former ice limits constrained by cosmogenic dating or offshore sediment sequences) around Antarctica. We review the progress towards building a pan-Antarctic age–depth model from these data by first introducing the relevant RES datasets that have been acquired across Antarctica over the last 6 decades (focussing specifically on those that detected internal-reflecting horizons) and outlining the processing steps typically undertaken to visualise, trace and date (by intersection with ice cores or modelling) the RES-imaged isochrones. We summarise the scientific applications for which Antarctica's internal architecture has been used to date and present a pathway to expanding Antarctic radiostratigraphy across the continent to provide a benchmark for a wider range of investigations: (1) identification of optimal sites for retrieving new ice-core palaeoclimate records targeting different periods; (2) reconstruction of surface mass balance on millennial or historical timescales; (3) estimation of basal melting and geothermal heat flux from radiostratigraphy and comprehensive mapping of basal-ice units to complement inferences from other geophysical and geological methods; (4) advancement of the knowledge of volcanic activity and fallout across Antarctica; and (5) refinement of numerical models that leverage radiostratigraphy to tune time-varying accumulation, basal melting and ice flow, firstly to reconstruct past behaviour and then to reduce uncertainties in projecting future ice-sheet behaviour.</p
Experimental observation of topological transition in optical multimeron
No full textTopological textures exemplify matter and energy configuration governed by inherent physical laws. Recently skyrmion-like textures possessing fixed topology have been realized in diverse optical systems, ranging from evanescent fields, structured media to free space. Magnetic skyrmions, with their tunable topological features via external stimuli, hold promise for innovative devices, such as memory, logic gates, and neuromorphic computing. However, achieving dynamic control over non-trivial topological transitions in optical counterparts remains elusive, limiting their application in areas such as super-resolution imaging and information processing. In this study, a platform is proposed and demonstrated for producing optical spin multimerons via multiple plasmonic vortex interference. More importantly, to manipulate the multimeron topology is managed by tuning the topology of the incident light, experimentally observing a topological transition with a nontrivial swap of the skyrmion number. This work broadens the scope of optical topological textures or quasiparticles, providing insights into their formation and transition mechanisms.</p