Animal affect, welfare, and decision-making: a computational modelling approach

People care about animal welfare because they assume that non-human animals including mammals and birds are able to experience negative emotions and hence to suffer. In order to monitor welfare and how it is affected by housing and husbandry, we therefore need accurate indicators of animal emotion. Because we can't be sure what other animals feel, and even whether they have the capacity for conscious experiences, our measures are necessarily indirect. Nevertheless, animal welfare scientists, neuroscientists and others are working to develop new and better indicators.

An area of growing interest is the use of cognitive markers of emotion, for example how different affective states alter the way in which decisions are made. We have developed a measure based on human psychology findings that people in negative affective states make more negative or pessimistic judgements about ambiguous events than happier people. Over 100 studies in a range of species have now been published using the 'judgement bias' (JB) test, and a meta-analysis that we are conducting indicates that, like humans, animals in more negative states exhibit more 'pessimistic' judgement biases. However, it also detects considerable variation in study findings. One important reason for this, which is receiving increasing attention in human cognitive neuroscience, is that emotional states influence a variety of hidden underlying decision processes which in turn determine the actual decisions made.

Such processes can be revealed by computational modelling of data from human decision-making tasks and include not only a person's expectations of good or bad decision outcomes, but also how they value these outcomes, how well they learn about changes in outcomes, and how strongly they adhere to what they have learnt when making a new decision. Anxious people, for example, appear to upgrade the anticipated unpleasantness of negative outcomes whilst depressed people downgrade the expected value of rewarding outcomes. Computational modelling techniques thus allow us to reveal hidden decision processes, evaluate how they are altered by emotions, and hence shed light on the complex links between affect and the actual decisions that we observe.

We will develop and implement a computational modelling approach to identify hidden decision-processes in animals, and how these are influenced by affective states. We will use a variant of our JB task that allows appropriate modelling, and a more naturalistic 'risky choice' test that doesn't require the pre-training necessary in JB tasks and hence is quicker to implement. We will induce both short- and longer-term positive and negative affective states using standardised manipulations. Computational modelling will then be employed to investigate how these influence underlying decision-making processes such as those previously identified in human studies.

Our work will introduce a novel computational modelling approach to the study of animal affect and welfare. It will identify new decision-making markers of affective states, including those that are replicable across decision tasks and hence particularly robust and reliable. It will also provide a deeper fundamental understanding of links between affect and decision-making processes in animals, and how similar these are to those observed in humans. This will indicate evolutionary similarities across species, and the potential for developing novel and translatable cognitive measures of affective state usable in animal welfare science and other fields. In order to open up the approach to other researchers, we will develop a Matlab toolbox, complete with code, examples, and documentation, that others can use to implement computational methods using similarly-designed JB tasks and our new risky choice task. In this way we hope to drive a step change in analytical methods, theoretical understanding, and new measurement tools that will advance animal welfare science.

An animal's affective state is a key determinant of its welfare. Accurate proxy measures of animal affect are thus required in animal welfare science and other fields. We have developed a 'judgement bias' (JB) measure of animal affect based on findings that people in negative affective states make more negative judgements about ambiguous events than happier people. A meta-analysis drawing from more than 100 JB studies indicates that this is also the case in animals. However, there is high heterogeneity in study findings. An important reason for this is that affective states may influence a variety of hidden decision processes which in turn determine actual decisions made. These include expectation and valuation of decision outcomes, rates of learning about changes in outcomes, and response stochasticity. Computational modelling of human decision task data shows that different affective states influence these processes in different ways. Changes in these processes are thus valuable new markers of affective states and disorders.

This approach has hardly been used in animal affect research, so our aim in this project is to develop computational modelling using our JB task, and a quicker to implement 'risky choice' task, to identify hidden decision-processes in animals and how these are influenced by standardised manipulations of affective state. We will thus be able to identify new decision-markers of animal affect; provide a deeper fundamental understanding of links between affect and decision-making processes in animals; show parallels with, and hence possible evolutionary similarity and translatability to, human findings; open up the approach to other researchers by creating a Matlab Toolbox complete with code, examples, and documentation to apply computational methods to JB tasks and our new risky choice task. In this way we will drive a step change in analytical methods, theoretical understanding, and new measurement tools that advance animal welfare science.

The proposed work will generate a new computational modelling approach in animal welfare science that can identify novel markers of animal affect and welfare, and new naturalistic and more rapidly implemented decision-tasks to study these. It will also establish similarities and differences between humans and laboratory animals in affect - decision-making links. In addition, the approach will significantly increase the amount of information gleaned from the sorts of experimental study of animal decision-making used here, and hence offer opportunities for decreasing the number of experiments required to address particular questions. These outcomes should be beneficial to a number of stakeholder groups in a variety of ways (benefits for other researchers are indicated in the Academic Beneficiaries section).

Laboratory animal technician community: The development of new markers of animal affect which have the potential to indicate the causes of positive or negative states (e.g. lack of reward; surfeit of punishment) and hence appropriate corrective measures, together with specific findings relating to the effects of treatments such as rat handling methods, will be of significance for animal caretakers seeking to better monitor and improve lab animal welfare.

Regulatory, funding and charitable bodies (e.g. Home Office Animals in Science Committee, NC3Rs, RSPCA Research Animals Department): Development of methods that allow more information to be extracted from animal experiments will be beneficial to stakeholders who are interested in ways of decreasing animal use in research. Computational modelling may thus provide a novel tool to aid 3Rs Reduction in animal numbers by reducing the number of experiments needed to acquire a set amount of information. These stakeholders will also be interested in any new measures that can be used to assess animal affect and welfare.

Industry - behavioural research equipment manufacturers: Companies that are developing new platforms for automating animal behaviour research such as Zantiks Ltd (https://zantiks.com), with whom we are in close contact, are interested in developing and modifying their equipment to assess animal welfare, and to run new cognitive and decision-making tasks. Our approach and findings will be of direct relevance in offering the opportunity to integrate computational modelling code alongside associated decision-making task programmes into a package that can be made available with their hardware.

General public: The public is inherently interested in animal behaviour, emotions and welfare. We will use this research area as an accessible and engaging context in which to introduce information about scientific methods and STEM research to the public, including school children. In particular, we will promote the value of learning to code as an important modern skill in all sorts of scientific areas, including animal welfare research. Our specific research findings on how emotions influence decisions and underlying decision-processes in both humans and animals, and how this knowledge can be used to assess animal welfare should also be of widespread general interest.

Research staff on the project: Our project combines whole animal experimental biology methods and techniques with sophisticated computational modelling approaches. Researchers on the project will thus acquire and develop quantitative skills that complement their core life science skills and hence broaden their knowledge, know-how and career prospects.