Keynotes | K&Ts | GACs | Talks | Posters | Search

Poster F in Poster Session F: Thursday, August 6, 1:45 – 3:30 pm, Kimmel Center, Shorin & Rosenthal Rooms

Adversarial construction for efficient experiment design in large task-spaces

Prakhar Godara¹, Frederick Callaway², Marcelo G Mattar¹; ¹New York University, ²Princeton University

Presenter: Prakhar Godara

A central obstacle to task-general cognitive modeling is experiment design: if behavior is studied over a large task space, which tasks should be run to reveal the underlying algorithm most efficiently? We propose adversarial construction (AC), an iterative procedure that fits a behavioral model to the data collected so far and then selects the next task to maximize the model's regret relative to a normative reference. We evaluate AC in a binary sequence-prediction task family generated by two-state hidden Markov models, using human data from 1,400 participants across 14 environments. AC converges within roughly six iterations to a compact set of diagnostic task types, including deterministic alternation, stochastic alternation, sticky regimes, and oddball structure. Models trained on AC-selected data show lower worst-case generalization error than models trained on randomly sampled tasks at every matched dataset size, despite using less data overall. AC-trained models also better capture held-out trial-level choice dynamics and acquire internal representations aligned with bigram statistics, the minimal representation sufficient for this task family. These results suggest that machine learning tools can be exploited for efficient task-space experiment design and AC provides a promising route toward more general computational theories of cognition.

Topic Area: Methods, Tools, Theory & Neural Coding