Probabilistic ODE Solvers via Bayesian Filtering and Smoothing
Author nathanaelbosch
44 Stars
Updated Last
2 Years Ago
Started In
October 2020


Stable Dev Build Status Coverage

ProbNumDiffEq.jl provides probabilistic ODE solvers for the DifferentialEquations.jl ecosystem.

Fitzhugh-Nagumo Solve Animation

The field of Probabilistic Numerics aims to quantify numerical uncertainty arising from finite computational resources. This package implements ODE filters, a class of probabilistic numerical methods for solving ordinary differential equations. By casting the solution of ODEs as a problem of Bayesian inference, they solve ODEs with methods of Bayesian filtering and smoothing. As a result, the solvers return a posterior probability distribution over ODE solutions and provide estimates of their own numerical approximation error.


The package can be installed directly with the Julia package manager:

] add ProbNumDiffEq


using ProbNumDiffEq

# ODE definition as in DifferentialEquations.jl
function fitz(u, p, t)
    a, b, c = p
    return [c*(u[1] - u[1]^3/3 + u[2])
            -(1/c)*(u[1] -  a - b*u[2])]
u0 = [-1.0; 1.0]
tspan = (0., 20.)
p = (0.2,0.2,3.0)
prob = ODEProblem(fitz, u0, tspan, p)

# Solve the ODE with a probabilistic numerical solver: EK0
sol = solve(prob, EK0(order=1), abstol=1e-1, reltol=1e-2)

# Plot the solution
using Plots
plot(sol, fillalpha=0.15)

# Sample from the solution and plot the samples
samples = ProbNumDiffEq.sample(sol, 100)
for i in 1:100
    plot!(sol.t, samples[:, :, i], color=[1 2], label="", linewidth=0.1)

Fitzhugh-Nagumo Solution