This is the (julia) code for the publication:
V. Liakoni *, A. Modirshanechi *, J. Brea, and W. Gerstner Learning in Volatile Environments with the Bayes Factor Surprise, Neural Computation 33, 269–340 (2021)
* V.L. and A.M. made equal contribution to this article.
Contact: vasiliki.liakoni@epfl.ch, alireza.modirshanechi@epfl.ch
Dependencies:
- Mac or Linux
- Julia (1.2)
Navigate into the NonstationaryModelLearning/src folder.
Open a julia terminal, press "]" to enter the package management mode and type
(v1.2) pkg> activate ..
(NonstationaryModelLearning) pkg> instantiate
All (julia) packages and dependencies will be installed automatically within this environment.
julia> using NonstationaryModelLearning
To run the main simulations of the paper, type
julia> NonstationaryModelLearning.runner_compare()
Note that the simulations would take long, so the settings are different from the ones used in the paper.
Uncomment the marked lines in the run_comparisons.jl file to replicate the paper's simulations.
Parameters of the environments and of the learning algorithms can also be changed in the same file.
Otherwise, a good place to start with the learning algorithms and the environments is the examples.jl file.
The naming convention of the algorithms are same as used in the paper,
apart from the Categorical task (section 2.3.2) which is called "Multinomial" in the code.
- gaussiantask: environment, learning algorithms and optimization functions for the Gaussian estimation task (section 2.3.1)
- multinomialtask: environment, learning algorithms and optimization functions for the Categorical estimation task (section 2.3.2)
- processandplot: results processing and plotting utilities.
- experiment: simulations for experimental predictions (section 2.4)
- parameterDicts: dictionaries with the optimized parameter values for different environments.
Please refer to src/readme.txt for more information.
- doc/data_final: final results used for the paper's figures.
There are some differences in naming conventions there, and we did not change them so that results are "as is" at publication time. Please refer to doc/data_final/readme.txt for more information. - doc/figs: latex source code to reproduce the paper's figures.