Skip to content

Parametrization and Initialization

There are many different ways in which a QAOA circuit can be parametrized, and even more the ways in which these parameters can be initialized.

Basic Parametrizations

We currently offer 7 different parametrizations for QAOA. They fall broadly into three categories:

  • The Standard classes are parametrization that have the \(\gamma\) 's and \(\beta\) 's as free parameters, as defined in the seminal paper by Farhi et al. [1].
  • The Fourier classes have the discrete cosine and sine transforms of the \(\gamma\)'s respective \(\beta\)'s as free parameters, as proposed by Zhou et al. [2].
  • The Annealing class is based on the idea of QAOA being a form of discretized, adiabatic annealing. Here the function values \(s(t_i)\) at equally spaced times \(t_i\) are the free parameters.

Except for the Annealing parameters, each class also comes in three levels of detail:

  • StandardParams and FourierParams offer the \(\gamma\) 's and \(\beta\) 's as proposed in references given above.
  • StandardWithBiasParams and FourierWithBiasParams allow for extra \(\gamma\)'s for possible single-qubit bias terms and their discrete sine transform, respectively.
  • ExtendedParams and FourierExtendedParams offer full control by having a separate set of rotation angles for each term in the cost and mixer Hamiltonians, respective having a seperate set of Fourier coefficients for each term.

Basic initializations

Before starting the algorithm, the parameters need to be initialized! That is, we need to give them some initial value. The choice of this value is arbitrary, and the act of coming up with good initial parameters is an ard on its own!

OpenQAOA supports three different initialization strategies:

  • rand (for random) sets the values of \(\beta\) and \(\gamma\) uniformly at random in the range \([0, \pi]\).
  • ramp create evenly spaced-time layers at the centers of p intervals, setting \(\beta\) and \(\gamma\) according to a linear ramp schedule. If time is not specified, the annealing time is set to \(0.7p\). For example, this specifies (\(\beta\), \(\gamma\)) = (0.35, 0.35) as the optimization starting point in a single-layer QAOA.
  • custom allows the user to initialize the angles with custom values.

The above can be combined to create a Variational Parameter Class with the Standard Parametrisation and Ramp Initializations

from openqaoa.qaoa_components.variational_parameters.variational_params_factory import create_qaoa_variational_params
create_qaoa_variational_params(qaoa_circuit_params=qaoa_circuit_params,
                                params_type='standard',
                                init_type='ramp')
or with the Extended Parametrisation and Random Initializations 
from openqaoa.qaoa_components.variational_parameters.variational_params_factory import create_qaoa_variational_params
create_qaoa_variational_params(qaoa_circuit_params=qaoa_circuit_params,
                                params_type='extended',
                                init_type='rand')

or with 

from openqaoa.qaoa_components.variational_parameters.variational_params_factory import create_qaoa_variational_params
create_qaoa_variational_params(qaoa_circuit_params=qaoa_circuit_params,
                                params_type='fourier',
                                q=1,
                                init_type='custom',
                                variational_params_dict={
                                    "betas":[0.26],
                                    "gammas":[0.42]
                                    }
                                )

Note that when using the fourier type, we need to specify the q value! For more details check out the fourier-parametrization page.

Using custom

When using the custom initialization strategy, we have to provide the appropriate number variational parameters in the form of a dictionary, variational_params_dict. The keys of the this dictionary are specific for the chose parametrization and the values depend on the number of values, p. Examples for how to use custom can be found in the separate pages.

References

  1. E. Farhi, J. Goldstone, S. Gutmann (2014). A Quantum Approximate Optimization Algorithm, https://arxiv.org/abs/1411.4028.
  2. L. Zhou, S. Wang, S. Choi, H. Pichler, M. D. Lukin, Phys.Rev.X 10, 021067 (2020) Quantum Approximate Optimization Algorithm: Performance, Mechanism, and Implementation on Near-Term Devices, https://arxiv.org/abs/1812.01041.