AnimalBehavior.jl

Generative models of animal behavior rely on the same global structure :

• They are defined by a set of latent variables (either fixed parameters or evolving variables)
• Those latent variables evolve as a function of external observations by an evolution function
• Actions are generated by sampling from distributions defined by an observation function of the latent variables

AnimalBehavior.jl takes advantage of this common structure to wrap some functionnalities of the Turing langage for dynamic probabilistic programming, in order to simulate and fit behavioral models with a minimal set of specifications from the user.

First, you need to create a DynamicPPL model using the @model macro, that returns all the latent variables of your model as a NamedTuple :

@model Qlearning(na, ns) = begin
    α ~ Beta()
    logβ ~ Normal(1,1)

    return (α=α, β=exp(logβ), Values = fill(1/na,na,ns))
end

MyModel = Qlearning(2,1)

Latent variables can be sampled from a prior distribution, and/or transformed by any arbitrary function

Then you have to define an evolution and an observation functions with the macros @evolutionand @observationrespectively with the following syntax :

@evolution MyModel begin 
        Values[a,s] += α * (r - Values[a,s]) # or : delta_rule!(s, a, r, Values, α)
    end

@observation MyModel begin
        Categorical(softmax(β * @views(Values[:,s])))
    end

The expression in the begin end statement can use the reserved variables names s, a and r for the current state, action and feedback respectively, and/or any latent variable defined earlier. Moreover, the observation function must return a Distribution from the Distributions.jl package.

# Simulation of a probabilistic reversal task
function pr_feedback(history) # Reverse the correct response every 20 trials
    correct = mod(length(history)/20, 2) < 1 ? 1 : 2
    return rand() < 0.9 ? history[end].a == correct : history[end].a ≠ correct 
end

sim = simulate(MyModel; feedback=pr_feedback);

simulate returns a Simulation structure with fields data and latent.

The package re-export the sample function to return a Chains object, with the following syntax :

sample(model, data, args...; kwargs...)

e.g. :

chn = sample(MyModel, sim.data, NUTS(), 1000)

WIP