Shout-out to my awesome co-authors - Reut Apel, Eyal Ert, Moshe Tennenholtz, David Bourgin, Josh C Peterson, Daniel Reichman, Tom Griffiths, Stuart J Russel, Even C Carter, James F Cavanagh, and Ido Erev.
Much more analysis in the paper, e.g., similar hybrid models predict risky individual choice and even strategic choice.
We believe the same approach may help in much more complex domains (stay tuned).
Thx for reading and sharing!
Full paper: rdcu.be/ew8o0
9/x
And best of all, BEAST-GB, trained on data from some experimental contexts predicts behavior in different experimental contexts, outperforming even direct empirical generalization.
8/x
In fact, on another (3rd) large dataset where BEAST miserably fails, the ML layer saves the day – BEAST-GB outperforms >50 competing models.
7/x
BEAST-GB is so accurate, we use the gaps between its predictions and BEAST's to enhance BEAST itself.
The GB algorithm detects idiosyncratic patterns in specific contexts as well as general patterns: task structures where BEAST's mechanisms are more or less active.
6/x
What if we have more data?
BEAST-GB, trained on minimal data, outperformed deep neural nets trained on 50x more data.
Using all data, BEAST-GB hit near-ceiling prediction: closing 96% of the gap between a perfect (hypothetical) model and random guessing.
5/x
BEAST? Not Prospect Theory?
Yes, BEAST (psycnet.apa.org/record/2017-...) – a behavioral model that won a previous competition. It assumes people mentally sample outcomes and choose accordingly.
Using it to integrate theory is much better than using classical models.
4/x
BEAST-GB integrates the behavioral logic of BEAST (an interpretable cognitive model) as features into a Gradient Boosting (GB) algorithm.
Analysis shows the combination is key: The behavioral BEAST features capture people’s sensitivities, and GB tunes them by context
3/x
In 2017, we launched CPC18, an open competition to predict human decisions under risk, under ambiguity, and from experience.
Computational models predicted hidden test data of people’s choices between lotteries with and w/o feedback.
Winner: BEAST-GB, a behavioral-ML hybrid
2/x
Proud & excited to share, 8 years(!) after starting this work.
🚨New paper in Nature Human Behaviour🚨
BEAST-GB: a hybrid model predicting human choice at near-ceiling levels.
It uses small data, beats behavioral & AI models, generalizes across contexts, and even explains human choice!
1/x
