Comparing likelihood-based and likelihood-free approaches to fitting and comparing models of intertemporal choice

Area of Science:

• Cognitive Science
• Computational Neuroscience
• Machine Learning

Background:

• Machine learning (ML) is increasingly used for cognitive model fitting and comparison, particularly for models lacking tractable likelihoods.
• Comparing ML approaches to traditional likelihood-based methods is crucial for understanding ML's utility in developing new cognitive models and theories.

Purpose of the Study:

• To systematically benchmark neural network (NN) approaches against likelihood-based methods for cognitive model fitting and comparison.
• To apply these methods to intertemporal choice modeling using data from individuals with substance use problems.

Main Methods:

• Benchmarking NN approaches against traditional likelihood-based methods.
• Applying both approaches to intertemporal choice data.
• Exploring extensions of NN approaches using recurrent and dropout layers for fitting complex data and posterior sampling.

Main Results:

• NN and Bayesian methods demonstrated convergence in inferring latent processes and substance use outcomes.
• Classification networks significantly outperformed likelihood-based metrics in model comparison.
• NNs are suitable for fast parameter estimation, posterior sampling, large datasets, and model comparison.

Conclusions:

• NNs offer advantages for specific modeling tasks like parameter estimation and comparison, especially with large datasets.
• Bayesian Markov Chain Monte Carlo (MCMC) methods remain preferable for smaller datasets with complex experimental designs.
• This study highlights the complementary roles of ML and traditional methods in advancing cognitive modeling.