Excellent question. Not sure, it looks like there might be circadian effect on Ghrelin levels? This might also be the reason for a sort of ceiling effect wrt the effect of sleep deprivation on Ghrelin.

It was a big team effort, led by Ella Brands, so congrats to everyone - great to see this out.

Results confirm that tyrosine supplementation affects reinforcement learning and decision-making functions that are thought to depend on catecholamine systems.

Performance was improved under Tyrosine, and computational modeling showed that this was due to an increase in value-dependent exploitation, and a reduction in exploration.

Here we studied reinforcement learning in n=63 participants (gender-balanced, within-subjects, preregistered) using a standard tyrosine dosage of 2g vs. placebo.

Tyrosine supplementation may improve cognitive performance under some conditions, but well-powered studies on processes thought to be under catecholamine control are sparse.

High capacity RNNs still deviated from human learners in several aspects, but our results confirm that the degree to which RNN and human behavior align substantially depends on network capacity.

However, increasing RNN capacity increasingly increased the fit of a computational model with a directed exploration term, compared to a model without. Following a more than 10-fold increase in capacity, the majority of networks showed evidence for directed exploration similar to human learners.

To understand how RNNs solve the task, we analyzed their behavior using a range of reinforcement learning models. RNNs with low capacity showed no evidence for directed exploration, in line with our earlier work: link.springer.com/article/10.1...

In such tasks, human learners often use directed exploration, that is, they tend to sample options with uncertain values for information gain.

We studied how recurrent neural networks solve restless bandit problems, a standard reinforcement learning task frequently studied in cognitive and systems neuroscience.

And here is the access link: authors.elsevier.com/a/1kRfA4sIRv...

This is one of those pieces that really benefitted from super helpful Reviewer input, so thank you 🙏 !!!