A Connectionist Model of the Psychological Refractory Period and Cognitive Control

Repository for the Master's thesis "A Connectionist Model of the Psychological Refractory Period and Cognitive Control" (A. Cesmeci, Osnabrück University, 2026; supervisor: S. Musslick). The project replicates and extends the PRP simulation (Simulation Study 3) from:

Musslick, S., Saxe, A., Hoskin, A., Sagiv, Y., Reichman, D., Petri, G., & Cohen, J. (2020). On the rational boundedness of cognitive control: Shared versus separated representations.

A feedforward task network (shared vs. separated task representations, task cues as control signals, temporal persistence of task-set activity) is read out by a Leaky Competing Accumulator (LCA). Dual-task (PRP) trials contrast a functionally dependent pairing (B→A, shared representations) with an independent pairing (C→A, separated representations) across SOAs, under two strategic regimes: greedy Task-2 engagement (cue on at stimulus onset) and strategic engagement (reward-rate-optimal onset; Eq. 7 of the preprint).

Headline result

Task-2 head slopes are lawfully task-dependent (dependent ≈ −0.5 to −0.9, independent ≈ −0.4 to −0.5 under the strategic regime) and vary with persistence and strategy, jointly spanning the empirical head-slope range documented in the thesis' literature evaluation (−0.27 to −1.60) — while only the strategic regime reproduces the empirically observed flat Task-2 error profile. See output/plots/ensemble/thesis/ and model_results_table_100726.md.

Repository structure

prp_model/            Core package
  task_network.py       Feedforward task network (paper Fig. 13 architecture)
  nn_wrapper.py         Training / integration wrapper (persistence EMA)
  training_set.py       MATLAB-style single-task training patterns
  lca.py                LCA dynamics (Eq. 4): run_lca / run_lca_avg / run_lca_dist
  threshold_utils.py    Session-level threshold selection (dual-task SOA-mixture,
                        accuracy-constrained expected reward rate) + onset policy
  prp_simulator.py      Two-pass PRP trial and SOA sweep
  utils.py              Trial generation, checkpoint I/O, aggregation, units
scripts/
  run_prp_sweep.py      Ensemble PRP sweep CLI (training, thresholds, sweeps)
  plot_prp_sweep.py     Thesis figures: RT1+RT2 panels, error rates, onset delay
  plot_money_figure.py  Head slope × persistence × condition × strategy
  make_results_table.py Summary table (markdown + LaTeX) across runs
run_finals.sh           Provenance: exact thesis simulation runs (M/S/R)
run_extras.sh           Provenance: extended persistence range runs
prp_old/                Archived pre-refactor code (do not use)
output/                 Results JSONs, plots, tables (gitignored except tables)

Quick start

python -m venv .venv && source .venv/bin/activate
pip install -r requirements.txt

# Full pipeline for one configuration (trains 20 networks if checkpoints
# are missing, computes session-level thresholds, runs both conditions):
python -m scripts.run_prp_sweep --store_dir ensemble_ckpt_p09 --E 20 \
    --train_if_missing --persistence 0.65 --trials_per_soa 50 \
    --soa_start 1 --soa_end 20 --soa_step 2 --ITI 1.8 \
    --optimize_onset --workers 6 --plot

# All thesis runs (main / greedy / robustness):
bash run_finals.sh && bash run_extras.sh

# Figures and summary table from saved JSONs:
python -m scripts.plot_prp_sweep --json "output/results/E20_*_ITI18_*zcD*.json" --context talk
python -m scripts.plot_money_figure --json "output/results/E20_p0[5-8]*_ITI18_*zcD*.json" \
    --empirical_band -1.60 -0.27
python -m scripts.make_results_table --json "output/results/E20_*_ITI18_*zcD*.json"

Key configuration (thesis-final; full rationale in simulation_spec_090726.md)

Parameter	Value
LCA (Eq. 4)	dt/τ = 0.1, λ = 0.4, α = 0.2, β = 0.2, σ = 0.2, t0 = 0.15
Time calibration	1 step = 50 ms
ITI	1.8 s (Pashler & Johnston, 1989); robust across 0.5–4.0 s
Thresholds	per task role, session-level, dual-task SOA-mixture context; accuracy floors 0.99 (T1) / 0.95 (T2); z2 shared across conditions
Onset policy	Eq. 7 reward-rate-optimal onset (main); greedy engagement (secondary)
Ensemble	20 networks; deterministic seeding of training, thresholds, stimuli (yoked across conditions), and LCA noise

Reproducibility

All reported simulations regenerate from the committed shell scripts and seeds; result filenames encode the full configuration (E{n}_p{persistence}_..._ITI{iti}_s{noise}_zc{context}_af{floor}_fx{z1}_oo{onset}_od{window}). Trained checkpoints are not tracked; they regenerate deterministically via --train_if_missing (seed = network index). Development history and the rationale for every modeling decision are documented in troubleshoot.md.