VQC tutorial does not reproduce reported accuracy

[RyuAmakaze]

Environment

• Qiskit version 1.4.3:
• qiskit-algorithms 0.3.1:
• Qiskit Machine Learning version 0.8.3:
• Python version 3.13.5:
• Operating system Ubuntu 20.04.6 LTS:

What is happening?

I ran the version of the code provided in the VQC tutorial, but I was not able to reproduce the reported accuracy.

Official

Model                           | Test Score | Train Score
SVC, 4 features                 |       0.99 |       0.97
VQC, 4 features, RealAmplitudes |       0.85 |       0.87

Ours

Model                           | Test Score | Train Score
SVC, 4 features                 |       0.99 |       0.97
VQC, 4 features, RealAmplitudes |       0.65 |       0.60

How can we reproduce the issue?

The code I used for the experiment is available here:
Using the same code, the same parameters, the same random seed, and the same optimizer.

What should happen?

The version change should not significantly affect the accuracy.

Any suggestions?

No response

[OkuyanBoga]

Hi, it is expected to be different since initial points are set randomly by default internally if it is not provided:

qiskit-machine-learning/qiskit_machine_learning/algorithms/trainable_model.py

Lines 249 to 250 in a88432f

	elif self._initial_point is None:
	self._initial_point = algorithm_globals.random.random(self._neural_network.num_weights)