UnitaryHACK: Implement and test gradient free optimizer for circuit models

[Thommy257]

Task description

To enhance the performance of our quantum models, we are currently investigating new optimizers for lambeq. One promising candidate is the Rotosolve algorithm (Ostaszewski et al.). Rotosolve is a gradient-free optimizer that leverages the fact that the expectation value of circuit measurements are sinusoidal with respect to the circuit parameters. Hence, Rotosolve can find the optimal parameters without the need for calculating gradients, making it particularly well-suited for optimization in the context of variational quantum algorithms in the NISQ era.

Your task is to implement the Rotosolve algorithm in lambeq, using the Optimizer interface. The following steps can be used as a guideline:

1. Familiarise yourself with the theory behind Rotosolve by reading the paper by Ostaszewski et al.
2. Investigate other optimizers in lambeq, e.g. the SPSAOptimizer
3. Implement a new class called RotosolveOptimizer which inherits from Optimizer.
4. Write unit tests to validate the functionality and performance of the Rotosolve optimizer.
5. Provide clear documentation and examples on how to use the Rotosolve optimizer within lambeq.

Notes

• Rotosolve is designed to minimise the expectation values of Hamiltionian measurements. Therefore, it won't work with classical losses like the Binary Cross Entropy. Assume that we are only optimizing a small number of variational quantum circuits where the loss is a linear combination of measurement expectation values.
• Rotosolve is an optimizer that does not keep an internal state. Hence, there is no need to provide any functionality for the state_dict and load_state_dict methods.
• The Optimizer interface has an attribute called gradient. Although Rotosolve is gradient-free, you can use the attribute to store the parameter updates.

Resources

• "Structure optimization for parameterized quantum circuits", Ostaszewski et al.
• "Quantum circuit structure learning" PennyLane blog
• PennyLane implementation

Alternative

If you don't manage to implement and test the RotosolveOptimizer, you can also implement a different (gradient-free) optimizer of your choice. Some ideas are COBYLA, Nelder-Mead, Conjugate-Gradient...

[Thommy257]

Hi everyone, please use the following notebook to test your optimiser. Please add it to the PR after you've ran it using your own optimiser (replacing SPSA). The notebook will need to be removed before merging, but it's important to visualise the performance first.

https://drive.google.com/file/d/1AT8q_crTWUFJ2hS2bNSrnc5T8WrinSnv/view?usp=share_link

Use jax==0.4.6 and jaxlib==0.4.6. There is currently a problem with newer versions of jax.

[vivek-kumar9696]

Hi, is this issue still open. I have an implementation of CMA-ES optimizer. Is it of value for the project?

[Thommy257]

Hi @vivek-kumar9696, of course! Feel free to submit a PR, I'd be happy to have a look at it.

[RamAIbot]

Hi @Thommy257 I would like to work on adding Rotosolve algorithm for unitary hack.

[dimkart]

@RamAIbot Hi and thanks for your interest. Since there is already a PR open for Rotosolve (#93) , it might be better to work on some other gradient-free optimiser, see some alternatives in Issue #85. If you open a PR we would be happy to review.

[Gopal-Dahale]

Hi @Thommy257, how about designing an interface namely ScipyOptimizer with which we can use Scipy's methods for optimization (scipy.optimise.minimise)?

[dimkart]

This is now completed. Thank you all for your work!