For the past two years I have been developing an independent line of research exploring learning, decision-making and meta-cognition across risk domains. I am also pursuing a range of related projects, from the compositionality of value representations to food decision-making, with internal and external collaborators. 

Brief outlines of some of this work is provided below. The sections below reflect ongoing work (submitted or in preparation). All copyright belongs to the author. 

Generalisable and transferable learning of risky choice

What does scaling Ben Nevis, investing ones’ savings and deciding which route to take to work have in common? All involve deciding between multiple potential actions, actions which are more or less likely to succeed. Crucially, riskier actions tend to be more valuable. Investing in stocks, for example, is riskier than investing in bonds, but also involves larger potential gains. 

For decennia, researchers have probed how humans’ trade-off risk and reward with a canonical verbal paradigm, asking questions such as: “Do you prefer £100 with a .4 probability, or £200 with a .25 probability?” This research has shown – with remarkable consistency – that human risky choice is stereotyped and sub-optimal. 

Recently, however, the generality of the canonical findings has been called into question by two separate lines of research. One line has focussed on choices in which outcomes are not described – but must be learned. The other has focussed on low-level sensori-motor or perceptual decisions. The conclusions from this research is radically different – either that human risky choice is differently sub-optimal, or much closer to optimal. 

Such dissociations– across paradigms with different risk domains – imply that risk-reward trade-offs are inherently domain-specific, and that the Nobel prize winning  work is severely restricted in scope. Strong specificity is also implied in recent interventions to improve human decision-making. Nudge-type interventions are designed to improve highly specific decisions – often single decisions in highly specific domains (e.g., pension savings) – rather than decision-making in general. 

On the other hand, strong dissociations seem incompatible with the prevailing view that a key feature of human cognition is our ability to generalise from a limited set of learning stimuli, and to transfer that knowledge across narrow domains. This view is also evident in research on artificial intelligence (AI), where it is often assumed human-like ability to generalise and transfer knowledge is one of the key missing pieces in creating a general AI. 

Learning better decisions across risk domains - paradigm illustration. Participants took part 
in pre- and post-training sessions performing a pointing task and a decision-making task. 
Between pre- and post- participants were trained on one choice task only either with or 
without feedback. 

In a double-blind randomised-control trial, we ask whether people can tune their risk-reward trade-offs to make better choices – not just between two options, but between a large, potentially infinite, set of options. A positive answer would be first evidence of generalisable learning of risky choice. To address mechanism specificity, we trained participants in one risk domain only, and tested for transfer of learning to un-trained domains. Compared to the standard observational approach in both behavioural economics and neuroscience, the method of experimental manipulation of risk preferences allows for unparalleled causal inferences (in prep.). Whilst theoretical - this works raises the possibility of broad general-purpose interventions to improve human decision-making in real world contexts. 

The world in which we operate is highly complex. Part of the problem is that the things we learn about (e.g., how useful a tree is for making a bow) have many features (e.g., age, bark, straightness, hardness), each of which can take many values (e.g., hardness: very soft, soft, hard etc). A pre-historic ancestor, for example, trying to learn which trees make the best bows would be faced with a total of feature^attribute values to learn. For any non-trivial set of features and attribute values, learning the value of individual trees quickly becomes intractable (i.e., the curse of dimensionality). However, one might exploit the compositionality in the environment – and learn not the value of individual trees but learn the values of features and attributes –  to predict the value of any tree. 

Together with Chris Summerfield and Hamed Nili, I have been exploring how people may use the compositional structure of the environment to bootstrap their learning of composite stimuli from their knowledge of elemental stimuli – much like our ancestors might have learned that a particular hardness of wood and a particular age of a tree in combination makes a really good bow. Using well matched visual stimuli, we taught participants the value of elemental stimuli, which they then could use to bootstrap their learning of compositional stimuli. We were especially interested in how the brain represents compositional and elemental value codes, and the project is designed around exploring value representation in the brain using fMRI and representational similarity analysis (RSA). 

The decisions we make about when and what to eat affects us both as individuals and as members of society. Sub-optimal decisions can lead to a host of adverse individual-level outcomes (e.g., obesity, cancer, cardiovascular events), which feed through to the societal level as an increased burden on families and healthcare systems. 

Together with Jeff Brunstrom and Peter Rogers at the University of Bristol, I have been exploring how one of the key risks– obesity – is associated with individual differences in how we make decisions about what to eat (in prep.). 

Using a large community sample, with each participant making several hundreds of simulated food choices, we developed a model that accounts 84% of participants choices. Jointly with known predictors of fat mass, individual differences in the modelled food-choice dimensions accounted for 42% of the variability in fat mass.

The observed level of prediction is far superior to that achieved by established predictors (e.g., socio-economic status) of fat mass alone (28% of the variability in fat mass). Notably, the work highlights specific dimensions of food choice that may be targeted by future interventions to optimise individual-level food choices. 

Example judgment tasks and choice task for exploring how individual differences in 
food choice strategies predict fat mass. Participants first rated foods on a number of
subjective food dimensions (e.g., perceived healthiness). Later they made decisions
between pairs of food options. 

Building on this work, we are exploring the neural substrate of the food choice model and the underlying different food dimensions, with initial results showing good correspondence between model-predicted food value and known value regions in the human brain (e.g., vmPFC). 

Canonical decision-making research has focused on so called decisions from description: “Would you prefer £10 with a.1 probability, or £20 with a .05 probability”. However, when we make decisions from experience – that is, when we have to learn the outcome distributions of choice options through repeated sampling – our decisions look radically different. 

Ben Newell, Chris Donkin and Jared Hotaling at the University of New South Wales, Casismir Ludwig at the University of Bristol, and I have been looking at an hitherto unexplored factor which may drive such dissociations (in prep.). Using Bayesian hierarchical modelling we find that once this unexplored factor is manipulated – experienced-based decisions start to look more like decisions from description. 

This work thus questions the extent to which decisions-from-experience are different because they are from ‘experience’, or because they invoke specific computational mechanisms for solving the task at hand.    

Developmental trajectories of risky choice

Together with colleagues at Birkbeck (Denis Mareschal, Ulrike Hahn) and UCL (Tessa Dekker, Imogen Large) I have been developing a paradigm based on Jarvstad et al. (2013,PNAS) to explore how developmental trajectories of risky choice might emerge across different risk domains. In a first large study, we explored how decision-making across 4 age groups (6-7, 8-10, 9-11, and adults) finding signs of both convergence and divergence across domains.