Periodic Reporting for period 2 - SYNERGY (Developing new behavioural models at the intersection of psychology, econometrics and machine learning)
Période du rapport: 2023-08-01 au 2025-01-31
Predictions of human behaviour underpin many important decisions by government and industry. They form a key input to cost benefit analyses for large infrastructure schemes as well as predictions of the impact of new policy measures. These predictions are made using advanced mathematical structures that replicate human decision making and predict changes therein in future hypothetical settings. The need for predictions is ubiquitous to the point that it is hard to think of a single field where models do not form a core part of the analytic toolkit.
While models have been used routinely to predict demand across different fields for several decades, their true potential has been limited by a failure to build synergies across three key methodological disciplines. These three areas all have strengths and weaknesses.
Process modelling (PM) in psychology, neuro-economics, and cognate disciplines focusses on “how” decisions are made, developing models that closely resemble known theories of behaviour, with particular interest in neural correlates, temporal dynamics and incorporating psychometric measurements.
Machine learning (ML) focusses on the “outcome”, replicating current choices or making predictions of changes in behaviour as a function of changing inputs. Increases in computing power have led to a growing popularity of ML techniques.
Choice modelling (CM) is positioned in between PM and ML, relying on structures that make a convenient, but realistic, assumption about the behavioural process, i.e. they fix the “how”, and focus on inferring the “why”, establishing which characteristics influence choices, in what direction, and to what extent. Unlike in PM, the assumptions about the process are driven by trade-offs between behavioural flexibility and computational tractability, large-scale applicability, and policy-relevance. A majority of work relies on random utility maximisation (RUM), producing outputs suitable for economic analysis.
The SYNERGY project will develop a new paradigm for modelling human decision making that addresses the schism between these three key disciplines, leading to new Data-Driven Behavioural Models (DDBMs) that recognise the importance of i) capturing “how” decisions are made, ii) identifying “why” a specific decision is reached, and iii) accurately predicting the “outcome” of future decisions.
While models have been used routinely to predict demand across different fields for several decades, their true potential has been limited by a failure to build synergies across three key methodological disciplines. These three areas all have strengths and weaknesses.
Process modelling (PM) in psychology, neuro-economics, and cognate disciplines focusses on “how” decisions are made, developing models that closely resemble known theories of behaviour, with particular interest in neural correlates, temporal dynamics and incorporating psychometric measurements.
Machine learning (ML) focusses on the “outcome”, replicating current choices or making predictions of changes in behaviour as a function of changing inputs. Increases in computing power have led to a growing popularity of ML techniques.
Choice modelling (CM) is positioned in between PM and ML, relying on structures that make a convenient, but realistic, assumption about the behavioural process, i.e. they fix the “how”, and focus on inferring the “why”, establishing which characteristics influence choices, in what direction, and to what extent. Unlike in PM, the assumptions about the process are driven by trade-offs between behavioural flexibility and computational tractability, large-scale applicability, and policy-relevance. A majority of work relies on random utility maximisation (RUM), producing outputs suitable for economic analysis.
The SYNERGY project will develop a new paradigm for modelling human decision making that addresses the schism between these three key disciplines, leading to new Data-Driven Behavioural Models (DDBMs) that recognise the importance of i) capturing “how” decisions are made, ii) identifying “why” a specific decision is reached, and iii) accurately predicting the “outcome” of future decisions.
In the first half of the SYNERGY project, the focus has been on making contributions in the three separate disciplines covered by the project, while at the same time starting work on building bridges between disciplines.
In the context of process models, the focus has been on improving the applicability of psychological models to real-world applications, and carrying out comparisons between models from psychology and econometric models seeking to capture similar behavioural patterns. This has looked both at models covering learning of preferences and models capturing internal (mental) deliberation processes. At the same time, initial efforts have been made to apply reinforcement learning models to real world choice data.
In machine learning (ML), the work thus far has focussed on a number of very distinct areas, namely a) the improvement of behavioural realism of ML algorithms, b) the evaluation and benchmarking of the prediction performance of ML, the c) improvement of out of sample prediction, and d) increasing the richness of representation of behavioural diversity.
In choice modelling, the focus has been on further developing our abilities to exploit rich new data sources with high spatial resolution, and to improve convergence performance for complex models.
Finally, initial work has looked at ensemble techniques, building joint up models that bring together components from all three disciplines.
In the context of process models, the focus has been on improving the applicability of psychological models to real-world applications, and carrying out comparisons between models from psychology and econometric models seeking to capture similar behavioural patterns. This has looked both at models covering learning of preferences and models capturing internal (mental) deliberation processes. At the same time, initial efforts have been made to apply reinforcement learning models to real world choice data.
In machine learning (ML), the work thus far has focussed on a number of very distinct areas, namely a) the improvement of behavioural realism of ML algorithms, b) the evaluation and benchmarking of the prediction performance of ML, the c) improvement of out of sample prediction, and d) increasing the richness of representation of behavioural diversity.
In choice modelling, the focus has been on further developing our abilities to exploit rich new data sources with high spatial resolution, and to improve convergence performance for complex models.
Finally, initial work has looked at ensemble techniques, building joint up models that bring together components from all three disciplines.
A number of the developments that have taken place so far in the SYNERGY project already go beyond the state-of-the-art in the individual disciplines. This relates primarily to the developments in a machine learning context. First, a new approach has been developed that allows analysts to impose greater behavioural regularity on their models, thus helping their real-world applicability. Second, a model averaging approach has been developed that allows the full power of machine learning models to be exploited when the forecasting scenarios are in line with the training data, while ensuring robustness outside that area through bringing in evidence from more behaviourally grounded approaches.
Numerous further developments are anticipated for the remainder of the project. These relate to the first large scale implementation of the active inference principle from mathematical psychology, the development of choice models that integrate ideas on preference learning from psychology, and the development of machine learning approaches for exploring and capturing preference heterogeneity.
Numerous further developments are anticipated for the remainder of the project. These relate to the first large scale implementation of the active inference principle from mathematical psychology, the development of choice models that integrate ideas on preference learning from psychology, and the development of machine learning approaches for exploring and capturing preference heterogeneity.