Config.jl

A configuration manager for Julia.

Suppose you have a configuration file called "tdlambda.json", specifying a set of parameter settings for an experiment

{
    "save_path": "RandomWalk19/tdlambda",

    "experiment": {
        "class": "MarkovRewardProcess",
        "episodes": 10
    },

    "environment": {
        "class": "RandomWalk",
        "nstates": 19
    },

    "agent": {
        "class": "TDLambda",
        "gamma": 1.0,
        "metastep": [0.025, 0.05, 0.075, 0.1],
        "lambda": [0.0, 0.4, 0.8, 0.9]
    }
}

Note: any config file must have the "save_path" parameter. This specifies the directory in data/output which the data will be saved to. In this example, data will be saved to data/output/RandomWalk19/tdlambda

Initialize a manager to manage all the details of this config file

cfg = ConfigManager("tdlambda.json", @__DIR__)

The second argument specifies where the data directory should be setup. In this case, a directory data/ will be setup in the same directory as the experiment which ConfigManager was instantiated in.

Any lists of parameters in the lowest-level of the config can be swept over (in this case, cfg["agent"]["metastep"] and cfg["agent"]["lambda"]). The different parameter settings are linearized. In order to sweep all the parameters of this config, we can first check how many different parameters there are:

N = total_combinations(cfg)

Then, we need to parse each of the individual settings into a concrete parameterization:

for idx=1:N
    parse!(cfg, idx)
    
    run_some_experiment(cfg)
end

parse sets up the settings of a particular parameterization. After parsing the config, individual parameters can be accessed by indexing. For example, to this parameterization's "metastep" parameter, we can call cfg["agent"]["metastep"]. If you will be referencing certain nested parameters quite often, you can get the subconfiguration instead: subcfg = get_subconfig(cfg, "agent"). Then access parameters of the subconfig in the same way: subcfg["metastep"].

Note that parse! has a third argument -- the run number -- which defaults to 1. To do multiple runs of an experiment, you can therefore do:

for run=1:100
    for idx=1:N
        parse!(cfg, idx, run)
        
        run_some_experiment(cfg)
    end
end

The ConfigManager also takes care of saving data to the right place. Just collect whatever data you want during your experiment in a Dict() and pass it to the ConfigManager.

function experiment(cfg::ConfigManager)
    data = Dict()
    data["ValueError"] = Float64[]
    for i=1:1000
       push!(data["ValueError"], rand()) 
    end
    
    save(cfg, data)
end

Then load the data later using load(cfg) (where cfg is a parse!'d config).