Lotka-Volterra Bayesian Parameter Estimation Benchmarks

Vaibhav Dixit, Chris Rackauckas

Parameter Estimation of Lotka-Volterra Equation using DiffEqBayes.jl

using DiffEqBayes
using Distributions
using OrdinaryDiffEq, RecursiveArrayTools, ParameterizedFunctions, DiffEqUncertainity
ERROR: ArgumentError: Package DiffEqUncertainity not found in current path:
- Run `import Pkg; Pkg.add("DiffEqUncertainity")` to install the DiffEqUncertainity package.

using Plots
using DiffEqMonteCarlo
ERROR: ArgumentError: Package DiffEqMonteCarlo not found in current path:
- Run `import Pkg; Pkg.add("DiffEqMonteCarlo")` to install the DiffEqMonteCarlo package.


Initializing the problem

f = @ode_def_nohes LotkaVolterraTest begin
    dx = a*x - b*x*y
    dy = -c*y + d*x*y
end a b c d
ERROR: LoadError: UndefVarError: @ode_def_nohes not defined
in expression starting at none:1
u0 = [1.0,1.0]
tspan = (0.0,10.0)
p = [1.5,1.0,3.0,1,0]
5-element Array{Float64,1}:
prob = ODEProblem(f,u0,tspan,p)
ERROR: UndefVarError: f not defined
sol = solve(prob,Tsit5())
ERROR: UndefVarError: prob not defined

We take the solution data obtained and add noise to it to obtain data for using in the Bayesian Inference of the parameters

t = collect(range(1,stop=10,length=10))
sig = 0.49
data = convert(Array, VectorOfArray([(sol(t[i]) + sig*randn(2)) for i in 1:length(t)]))
ERROR: UndefVarError: sol not defined

Plots of the actual data and generated data

scatter(t, data[1,:], lab="#prey (data)")
ERROR: UndefVarError: data not defined
scatter!(t, data[2,:], lab="#predator (data)")
ERROR: UndefVarError: data not defined
ERROR: UndefVarError: sol not defined
priors = [Truncated(Normal(1.5,0.5),0.5,2.5),Truncated(Normal(1.2,0.5),0,2),Truncated(Normal(3.0,0.5),1,4),Truncated(Normal(1.0,0.5),0,2)]
4-element Array{Distributions.Truncated{Distributions.Normal{Float64},Distr
 Truncated(Distributions.Normal{Float64}(μ=1.5, σ=0.5), range=(0.5, 2.5))
 Truncated(Distributions.Normal{Float64}(μ=1.2, σ=0.5), range=(0.0, 2.0))
 Truncated(Distributions.Normal{Float64}(μ=3.0, σ=0.5), range=(1.0, 4.0))
 Truncated(Distributions.Normal{Float64}(μ=1.0, σ=0.5), range=(0.0, 2.0))

We use the uncertainity quantification plots to decide the tolerance to be passed.

cb = AdaptiveProbIntsUncertainty(5)
ERROR: UndefVarError: AdaptiveProbIntsUncertainty not defined
monte_prob = MonteCarloProblem(prob)
ERROR: UndefVarError: prob not defined
sim = solve(monte_prob,Tsit5(),num_monte=100,callback=cb,reltol=1e-5,abstol=1e-5)
ERROR: UndefVarError: cb not defined
ERROR: UndefVarError: sim not defined

Parameter estimation with Stan.jl backend

The solution converges for tolerance values lower than 1e-3, lower tolerance leads to better accuracy in result but is accompanied by longer warmup and sampling time, truncated normal priors are used for preventing Stan from stepping into negative values.

@time bayesian_result_stan = stan_inference(prob,t,data,priors;reltol=1e-5,abstol=1e-5,vars =(StanODEData(),InverseGamma(3,2)))
ERROR: UndefVarError: prob not defined

Plots of the chains generated to show convergence.

ERROR: UndefVarError: bayesian_result_stan not defined

Parameter estimation with Turing.jl backend

@time bayesian_result_turing = turing_inference(prob,Tsit5(),t,data,priors)
ERROR: UndefVarError: prob not defined

The chains seem to have not converged and require longer chains but there isn't anyway to pass warmup samples to Turing.jl's HMC sampler which has been used as the sampler in the implementation.

ERROR: UndefVarError: bayesian_result_turing not defined


Lotka-Volterra Equation is a "predator-prey" model, it models population of two species in which one is the predator (wolf) and the other is the prey (rabbit). It depicts a cyclic behaviour, which is also seen in its Uncertainity Quantification Plots. This behaviour makes it easy to estimate even at very high tolerance values (1e-3).

In case of Stan.jl backend we obtain quite accurate values by setting sufficiently low tolerance at 1e-5 and passing 500 warmup samples, because as evident from the plots as it didn't converge before it which was observed from multiple runs, which ensures both high accuracy within 1.84 minutes, 1.7 minutes for warmup sampling and 14 seconds for sampling. Decreasing the tolerance leads to more accurate results but at the cost of significant increase in time taken.

Turing.jl backend implementation doesn't seem to have converged, inability to pass warmup samples is one of the drawbacks, the results obtained are quite accurate and it recorded 33.33 seconds.

Using DynamicHMC.jl as the backend gives good accuracy and takes 197 seconds but the exploration of the domain seems to be more constrained as compared to other backends as evident from the plots which is due to the lower stepsize, this can be adjusted by passing the kwarg ϵ.