QAOALandscapes
QAOALandscapes is a Julia package for exactly simulating the QAOA algorithm for general p-spin problems with the goal of understanding and exploring the classical optimization landscape. We would like to understand why some of the heuristic initialization/optimization strategies out there work the way they do. Currently, three initialization/optimization strategies are implemented:
- Transition states.
- Interpolation (Interp) strategy.
- Fourier strategy.
In terms of optimization, we currently support all the methods available through Optim.jl. The computation of the cost function/energy gradient is done using the adjoint differentiation method from this very nice paper Efficient calculation of gradients in classical simulations of variational quantum algorithms.
Installation
To install QAOALandscapes, open the Julia REPL and run the following command:
] add https://github.com/RaimelMedina/QAOALandscapes.git
Usage
Here is an example of how to use QAOALandscapes to solve a MaxCut type problem:
using QAOALandscapes
using Graphs, Random
n = 10
d = 3 # for 3-regular random graphs
pmax = 10 # maximum circuit depth to explore
g = random_regular_graph(n, d) # 3-regular unweighted graph
prob = ClassicalProblem(Float64, g)
qaoa = QAOA(prob) # if the problem is Z2 symmetric then the algorithm will work in the correct Hilbert subspace
init_point, init_energy = getInitialParameter(qaoa) # obtain initial parameters at p=1
# Now choose a strategy
# For example, for transition states we have implemented the Greedy strategy
greedyData = greedyOptimize(qaoa, init_point, pmax);
# If you prefer INTERP strategy
interpData = interpOptimize(qaoa, init_point, pmax);
# Alternatively, you can also optimize a given set of parameters
# directly
initial_parameter = rand(20)
optimal_param, optimal_energy = optimizeParameters(qaoa, initial_parameter)
Documentation
Right now is not working but I’m working on that and it should be fixed soon.
GPU
As of now, I don’t know how to setup the package so that it automatically detects which GPU (if any) the user has. So for now, I recommend that after downloading the package you build it again by including the gpu.jl file and add the correspoding GPU package, i.e., CUDA.jl or Metal.jl. I plan to add support to AMD devices via KernelAbstractions.jl but that will come later
Warning
This is a work in progress and the code is very very rudimentary. I hope in this would be in a decent state to share sometime in the future.