RL for Psychometrics¶
A compact project that fuses reinforcement learning with psychometrics to infer interpretable individual ability from interaction trajectories.
Summary¶
We model human decision making as an MDP with a shared value function \(Q_\theta(s, a)\) and a per-participant inverse temperature \(\beta_j > 0\). Parameters are estimated jointly via a generalized EM procedure that balances behavioral likelihood and value-consistency regularization.
What’s in this project¶
- Methodology — task formulation, probabilistic policy, generalized EM, and regularization (λ_bell, λ_cql, β priors).
- Experiments — Peg Solitaire (4×4, 7×7), β recovery, convergence diagnostics, and Q-function evaluation.
Project goals¶
- Interpretability — connect RL dynamics to psychological constructs via \(\beta_j\).
- Identifiability — separate \(\beta\) and \(Q\) under joint estimation.
- Generalization — learn \(Q\) that transfers across participants and tasks.
Repository¶
Maintainer: Wenqian Xu