Given the fast evolution of working memory computational models and the methodology of modeling these years, a new toolbox that could estimate, assess and comprehensively compare the new models using the state-of-the-art methods should be helpful for the future studies. The Bayesian Modeling of Working Memory (BMW) toolbox is working on this issue by establishing a pipeline that integrates functions of model definition, model estimation, model comparison & selection. All users could either finish all the job by GUI or by recruiting Matlab scripts.
- Fit Models by Maximum Likelihood (MLE)/ Maximum A Posteriori (MAP)
- Available Methods for Model Assessment/Comparison
- Log Likelihood (LLH)
- Akaike Information Criterion (AIC)
- Modified Akaike Information Criterion (AICc)
- Bayesian Information Criterion (BIC)
- Deviance Information Criterion (DIC1/DIC2/DIC*)
- Watanabe-Akaike Information Criterion (WAIC1/WAIC2)
- Log Model Evidence/Marginal Likelihood (LME)
- Second-Level Random-Effect Bayesian Model Selection (RFX-BMS)
- Expected Model Frequency
- Exceedance Probability
- Available Models
- Continuous Recall Models
- Item Limit
- Standard Mixture
- Slots-plus-Averaging
- Equal Precision
- Variable Precision
- Variable Precision with Capacity
- Category-Only
- Category-Only (with Capacity)
- Model Variants: + Continuous Capacity/ResponseNoise/Bias/Swap/Bias Fluctuation/Precision Fluctuation/Categorical Encoding (Between-Item)/Categorical Encoding (Within-Item)
- Change Detection Models
- Fixed-Capacity (Single-Probe)
- Fixed-Capacity (Center-Probe)
- Fixed-Capacity (Whole-Display)
- Signal Detection
- Model Variants: + Lapse/Ensemble Encoding
- Custom Models
- The RW Reinforcement Model
- Von Mises Distribution
- Continuous Recall Models
- Available Built-In Optimization Algorithms
- Default: Differential Evolution Monte Carlo Markov Chain (DE-MCMC)
- Adaptive Metropolis-Hastings Monte Carlo Markov Chain (MH-MCMC)
See "Requirements" for other options
- Matlab (Best if >= 9.0/2016a, not sure if the lower version works)
- Other Optimization Algorithms (Optional)
- fmincon (sqp/interior point/active set), Optimization Toolbox (built-in toolbox in Matlab >= 9.0)
- Genetic Algorithm, Global Optimization Toolbox (built-in toolbox in Matlab >= 7.0)
- Simulated Annealing, Global Optimization Toolbox (built-in toolbox in Matlab >= 7.0)
- Mesh Adaptive Direct Search (MADS), Global Optimization Toolbox (built-in toolbox in Matlab >= 7.0)
- Bayesian Adaptive Direct Search (BADS), BADS Toolbox (http://github.com/lacerbi/bads)
3/24/2020, First full edition | Ma, Tianye & Dr. Ku, Yixuan
6/8/2020, Add categorical encoding strategies as model factors | Ma, Tianye & Dr. Ku, Yixuan
Sincere thanks to
Sizhu Han,
Ruyuan Dr. Zhang,
Kuangshi Zhao,
for their invaluable help!
Computational modeling could act as a powerful tool especially in understanding abstract issues such as "Do we have discrete capacity in working memory storage?" and "How is our working memory system organized?". And we suppose that models of best quality should on the one hand take a good balance of accuracy and complexity while on the other hand offer useful outputs. That is to say, we take the well-renowned George Box's view,
All models are wrong, but some are useful.
In order to make this toolbox a convenient tool and a good communicator of computational methods and working memory models for cognitive scientists, we'll always need feedback. So please feel free to start a Github issue or contact Ma, Tianye (mack_ma2018@outlook.com) for bug reports and any other kinds of advice.