SPICE Studies

weinhardt2026/studies/ holds self-contained per-study directories: each combines a task/data-loading module, a SPICE model config (spice_<study>.py), a hand-coded benchmark model, and a benchmarking_<study>.py file wiring generative benchmarking (see analyses.md). Every study directory typically has data/, params/, results/, and either a notebook or a <study>.py driver script.

Conventions and infrastructure shared across studies are described in training.md (model/config/estimator internals) and analyses.md (evaluation, morphing, coefficient statistics, generative comparison). This file covers what’s specific to each study: the task paradigm, population, and benchmark model.

---

Active Studies

braun2018 — Cognitive-control / task-switching effort

Participants choose which of two tasks to perform on each trial (df_choice='transcode') under a reward and switch/repeat structure. SPICE models separate reward, control, and fatigue value modules split by repeat vs. switch trials, plus a fatigue module driven by block number (control becomes more costly the longer it is exerted). Effort-based task selection and switch costs.

bruckner2025 — Helicopter / predictive-tracking task

Participants track a hidden target position (helicopter/coin drop) and adjust a “bucket” under changing volatility and stochasticity. Models changepoint detection and uncertainty-driven adaptive learning rates, plus an anchoring-bias correction module for bucket displacement between trials. Benchmarked against a RationalResourceModel (bounded-rational, resource-rational learning-rate model).

bustamante2023 — Patch-foraging task

Participants decide to harvest a depleting resource patch or exit to a new one. SPICE models reward/depletion tracking per patch state plus a “continuation” (stay/leave) module. Benchmarked against a Marginal Value Theorem model (MarginalValueTheoremModel) — the classic optimal-foraging economics baseline.

dezfouli2019 — Working-memory-mediated bandit learning

Bandit-style value-learning task modeled with working-memory buffers: reward/choice history over 3 lagged timesteps (t-1..t-3) feeds separate chosen/not-chosen value modules via spice.precoded.workingmemory. Benchmarked against a GQLModel (generalized Q-learning). Models multi-timescale reward learning with short-term dependence on recent reward/choice sequences.

eckstein2026 — Directed exploration in reward learning

Reward-learning bandit task modeling both value learning (per-environment/chosen/unchosen reward tracking with a running-mean baseline) and directed exploration (separate positive/negative value-difference “exploration” modules), plus choice perseveration and spatial-attention biases (adjacent/opposite action bias). Benchmarked against Castro2025Model and RWForgettingChoiceModel (Rescorla-Wagner with forgetting). Associated with the “MindRL Challenge 2026” (mindrl_challenge_2026.ipynb).

ganesh2024a — Confidence-weighted reward learning

Perceptual decision-making combined with reward learning: a perception_certainty module maps signed contrast difference to a confidence/certainty signal (sigmoid), gating separate chosen/unchosen reward-learning and choice-persistence value modules. Benchmarked against a BayesianModel. Demonstrates the items-vs-actions decoupling (see training.md).

groman2018 — Planned, not yet implemented

Reference material only (groman2018.pdf — Groman et al., Neuropsychopharmacology 2018, on reinforcement-based decision making in rats under chronic methamphetamine exposure). No groman2018.py, spice_groman2018.py, or benchmarking_groman2018.py exists yet; data/, params/, results/, figures/ are placeholders.

huang2026 — Collaborative visuospatial foraging

A collaborative tile game between a self participant and a partner. Models tile-value dynamics from information loss (decay when unvisited), information gain (decrement on visiting), recency of self/partner visits, choice stickiness, and movement perseverance. Benchmarked against a 7-parameter InformationForagingModel.

kolff2025 — Primate social behavior

Models a 5-action repertoire (action/grooming/gesture/scratching/waiting) as a function of a partner’s simultaneous behavior and dominance-rank difference (rank_diff), with separate per-dimension “partner influence” and “own-action transition” modules plus persistence (repeat/switch) modules. Additional inputs include per-individual dominance rank. Benchmarked against a ConditionalFrequencyModel.

weber2024 — Belief updating under volatility (angular state space)

A laser-shield tracking/belief-updating task (“catch”/”miss” a laser with a shield) under volatility and stochasticity: changepoint + relative-uncertainty learning-rate modules gate a circular belief-position update via sin/cos, plus a dynamic learning-rate state. Same predictive-inference family as bruckner2025 (Nassar-style changepoint/volatility belief updating), but with an angular (not linear) state space — evaluated with circular/angular MSE loss (clamped_angular_mse). Benchmarked against a ChangePointModel.

---

synthetic — Parameter Recovery Testbed

Not a real-data study. Contains benchmarking_qlearning.py and synthetic_<Np>p_<seed>_<idx>.csv datasets simulating Q-learning agents at varying population sizes (32/64/128/256/512 participants). Used to validate that SPICE recovers known ground-truth cognitive parameters/dynamics — see analysis_parameter_recovery.py in analyses.md.

---

archive — Inactive Studies

augustat2025, eckstein2022, rtify2024, sugawara2021, ting2023 — earlier or superseded study directories, kept for reference but not maintained against the current SPICE API.

---

Adding a New Study

1. Create weinhardt2026/studies/<study>/ with data/, params/, results/.
2. Write a data-loading module (get_dataset()) using csv_to_dataset() (see training.md).
3. Define spice_<study>.py: a BaseModel subclass + SpiceConfig following the precoded model pattern, applying the polynomial-amenable architecture guidelines.
4. Write a hand-coded benchmark model (for analysis_model_evaluation.py comparison) and an Environment<Study>(Env) + generate_behavior() wrapper (for generative benchmarking — see analyses.md).
5. Fit with SpiceEstimator, then run the typical analysis sequence.