This repository contains the code to run agent-based simulations of a word association game performed either individually or interactively by agents of varying diversity, whose semantic memories are noised versions of a skip-gram word2vec model.
- Run for multiple thresholds
- Tidy up code
The repository includes:
diversity_abm- Code for the mechanisms of the simulation;
baseline.ipynb- Notebook to run and analyse baseline performance (of the non-noised word2vec agent);
noise_and_pair.ipynb- Code to add varying levels of noise to the word2vec model and generate agents of different "diversity levels";
analysis.ipynbandsupplementary.ipynb- analysis of the effect of interaction and diversity on different metrics of creativity in the association behavior produced by agents/pairs.
A first version of this is published as CogSci proceedings: https://escholarship.org/uc/item/58v5d82w, cite as: Rocca, R., & Tylén, K., 2022, Cognitive diversity promotes collective creativity: an agent-based simulation, Proceedings of the Annual Meeting of the Cognitive Science Society, 44
A follow-up is in progress.
Current setup with n-back agents
- Strict: if agent does not find associations in n-back turns, trial is over;
- Flexible: if current agent does not find association in n-back turns, turn is handed over to the other agent
- Note, instead of nesting the n-back logic inside the flexible interaction logic, one could do the other way round;
- Shortest: for each seed, give word to the agent with shortest association. If no sub-threshold values are available, go one seed back and repeat.
Future steps
- Look at diversity with multiple possible strategies
- Invariant (semantic + phonological) + variable (episodic)
- Diversity as probability of picking one over the other?
- Implement pairs based on human data