Convention Formation and Transmission.

This repo implements a minimal cultural evolution sandbox inspired by Compositional Abstractions Tutorial.

The report of this project is here: Convention Formation and Transmission

Requirements

• Python 3.10+ (should work on 3.9+).
• Dependencies: python -m pip install -r requirements.txt

Project Layout

• data/task/task.json — tower tasks and their correct programs. Extracted from the original project above.

• data/task/lexicon.json — lexicon dict includes meaning, utterance and the associated program. Extracted from the original project above.

• config.py — hyperparameters and paths.

• task.py, lexicon.py — task/lexicon utilities.

• agents.py — DyadAgent class.

• population.py — population class.

• exp1_dyad_convention.py — single-dyad demo (round-level logs + curves).

• plot.py — heatmaps for Exp1.

• exp2_population_convention.py — paired vs mixed training mode for population.

• plot2.py — heatmaps for Exp2.

• exp3_generation_convention.py — multi-generation turnover experiment.

• plot3.py — heatmaps for Exp3.

Outputs:

• Plots → plots/
• JSON logs → results/

Usage

Exp1: Running a Single Dyad

python exp1_dyad_convention.py
python plot.py

Outputs:

• Plots under plots/exp1/
• JSON logs under results/exp1/

Exp2: Population Experiments(Paired vs Mixed)

python exp2_population_convention.py
python plot2.py

Outputs:

• Plots under plots/exp2/
• JSON logs under results/exp2/

Exp3: Turnover Across Generations

python exp3_generation_convention.py
python plot3.py

Outputs:

• Plots under plots/exp3/
• JSON logs under results/exp3/

Configuration

All global knobs live in config.py.

Parameter	What it controls (code)	If increased…	If decreased…
LENGTH_PRIOR_LAMBDA	Length bias (lexicon.py, agents.py)	Note: in lexicon.py the prior uses exp(-λ*(L-1)) (λ↑ favors shorter meanings/programs), but in agents.py the task→program initialization uses exp(+λ*(L-1)) (λ↑ favors longer programs).	Closer to uniform length preference (length trends become more data-driven).
FIDELITY	Strength of aligned meaning↔utterance pairs in the prior (lexicon.py)	More aligned prior → often higher early accuracy and less exploration.	Flatter prior → more exploration and slower convergence.
ALPHA_SPEAKER	Rationality in program choice (agents.py: sample_program_for_task)	Sharper softmax and more greedy.	More stochastic exploration.
TASK_COST_WEIGHT	Weight of program-length cost β (agents.py: sample_program_for_task)	Stronger preference for short programs.	More driven by belief probabilities.
BELIF_UPDATE	Belief update step size (agents.py: update_*)	More aggressive updates and faster changes.	More conservative updates.
CHOICE_TEMPERATURE	Sampling temperature τ (agents.py: _apply_temperature)	More random sample.	Closer to argmax.

Experiment defaults

Parameter	What it controls
ROUNDS_PER_DYAD	Rounds per generation.
N_DYADS	Number of dyads in the population (Exp2/Exp3).
NEWPOP_RATIO	Fraction of dyads replaced between generations (Exp3).
GEN_POP	Default number of generations (Exp3 default).
GLOBAL_SEED	Random seed.
EPSILON	Numerical stability constant for safety.