Hello,
I am trying to implement a custom variational quantum algorithm to solve the unweigthed Maximum Cut problem. It is inspired by the QAOA algorithm (in fact, it uses the same ansatz), but features some modifications, in turn inspired by this paper: https://arxiv.org/abs/2308.10383.
The problem that I’m running into is that, while trying to optimize my ansatz’s parameters (to minimize a specific cost function), using the Adam optimizer, through qml.AdamOptimizer, I notice that the parameters aren’t being updated. The output that I get, while training, looks like this (after 5 optimization steps):
Optimizing parameters...
Objective after step 1: 0.750000000000
Parameters after step 1: [[0.04851585 0.02871953]
[0.04399233 0.08008254]].
Objective after step 2: 0.750000000000
Parameters after step 2: [[0.04851585 0.02871953]
[0.04399233 0.08008254]].
Objective after step 3: 0.750000000000
Parameters after step 3: [[0.04851585 0.02871953]
[0.04399232 0.08008254]].
Objective after step 4: 0.750000000000
Parameters after step 4: [[0.04851585 0.02871953]
[0.04399232 0.08008254]].
Objective after step 5: 0.750000000000
Parameters after step 5: [[0.04851585 0.02871953]
[0.04399233 0.08008254]].
I suspect the problem might come from how I define the objective function, but I haven’t been able to resolve this. I also tried testing the hypothesis of my optimization landscape being just really flat, but I wasn’t succesful. My idea was to compute the gradients, using the compute_grad method, from qml.AdamOptimizer, but I was having trouble getting this to work, as it appears to require a positional kwargs argument that I don’t know what is supposed to be in my case (I was trying running compute_grad(objective, parameters), but that didn’t work, as the function required something for kwargs. On the other hand, setting kwargs = {} led to other errors.)
To fully reproduce these results, here’s the utilized code:
# Required imports
import pennylane as qml
from pennylane import numpy as np
import matplotlib.pyplot as plt
import itertools
import time # For timing
# 8-node graph:
graph = [(0, 1), (0, 2), (0, 6), (1, 2), (1, 6), \
(3, 2), (3, 4), (3, 5), (4, 5), (4, 7), \
(5, 7), (6, 7)]; n_nodes = 8
# Basis states lists, to be associated with each graph node
n_qubits, basis_states_lists = n_nodes, []
permutations = ["".join(seq) for seq in itertools.product("01", repeat = n_qubits - 1)]
for i in range(n_qubits):
individual_list = []
for perm in permutations: individual_list.append(perm[:i] + "1" + perm[i:])
basis_states_lists.append(individual_list)
# VQA implementation
def iQAQE_QAOA_Ansatz(graph, n_qubits = None, n_layers = None, device = 'default.qubit', shots = None,
parameters = None, B = None, basis_states_lists = None,
max_iter = 100, step_size = 0.99):
"""
Implements the I-QA/QE VQA, for the MaxCut problem. [Modified to use the QAOA ansatz!
Will only work if n_qubits = n_nodes.
Not protected against other cases!]
"""
# Prelimary definitions
def compute_nodes_probs(probabilities, basis_states_lists):
"""
Computes the probabilities associated with each node.
Args:
probabilities (list): The probabilities of each basis state.
basis_states_lists (list[list]): List of lists of basis states assigned to each graph node.
Returns:
list: The probabilities associated with each node.
"""
nodes_probs = []
for sublist in basis_states_lists:
node_prob = 0
for basis_state in sublist: node_prob += probabilities[int(basis_state, 2)]
nodes_probs.append(node_prob)
nodes_probs = [prob/sum(nodes_probs) for prob in nodes_probs]
return nodes_probs
assert(n_qubits is not None and n_layers is not None and parameters is not None and B is not None and basis_states_lists is not None), "n_qubits, n_layers, parameters, B and basis_states_lists cannot be None."
# Device setup
dev = qml.device(device, wires = n_qubits, shots = shots)
# Circuit setup
# Unitary operator 'U_C' (Cost/Problem) with parameters 'gamma'.
def U_C(gamma):
for edge in graph:
wire1, wire2 = edge[0], edge[1]
qml.CNOT(wires = [wire1, wire2])
qml.RZ(gamma, wires = wire2)
qml.CNOT(wires = [wire1, wire2])
# Unitary operator 'U_B' (Bias) with parameters 'beta'.
def U_B(beta):
for wire in range(n_qubits):
qml.RX(2 * beta, wires=wire)
# QAOA circuit ansatz
@qml.qnode(dev)
def circuit(gammas, betas):
for wire in range(n_qubits):
qml.Hadamard(wires = wire)
for i in range(n_layers):
U_C(gammas[i])
U_B(betas[i])
return qml.probs()
# Draw the circuit, for visualization
qml.drawer.use_style("pennylane") # Set the default style
print(f'Quantum circuit drawing: n_qubits = {n_qubits}, n_layers = {n_layers}.')
fig, ax = qml.draw_mpl(circuit, decimals=3)(parameters[0], parameters[1])
plt.show()
# Status message, before optimization
print(f"iQAQE level (# of layers): p = {n_layers}.")
# Define the cost function
def objective(params):
# Get the parameters
gammas, betas = params[0], params[1]
# First - Compute the probaility associated to each node, from the 'n_qubit' probability distribution
probs = circuit(gammas, betas); cost = 0
nodes_probs = compute_nodes_probs(probs, basis_states_lists)
# Second - Compute the cost function itself (From https://arxiv.org/abs/2308.10383)
for edge in graph:
# j and k are the nodes connected by the edge
# 0: j, 1: k
d_jk = np.abs(nodes_probs[edge[0]] - nodes_probs[edge[1]]); s_jk = nodes_probs[edge[0]] + nodes_probs[edge[1]]
edge_cost = (d_jk - 1/B)**2 + (s_jk - 1/B)**2
cost += edge_cost
return cost
# Initialize optimizer: Adagrad works well empirically. We use Adam, though.
print(f"Step size: {step_size}.")
opt = qml.AdamOptimizer(stepsize = step_size, beta1 = 0.9, beta2 = 0.99, eps = 1e-08)
# Optimize parameters in objective
start_time = time.time(); i, cost_vec, ar_vec = 0, [], [] # For timing
print("\nOptimizing parameters...")
while(True):
parameters, cost = opt.step_and_cost(objective, parameters); i += 1; cost_vec.append(cost)
print(f"Objective after step {i:4d}: {cost:.12f}")
print(f"Parameters after step {i:4d}: {parameters}.")
# Check maximum iterations
if i == max_iter:
print(f"Maximum number of iterations reached: max_iter = {max_iter}")
break
train_time = time.time() - start_time # Time taken for training
print("--- Training took %s seconds ---" % (train_time))
# Just take the exact distribution
probabilities = circuit(parameters[0], parameters[1])
# Computing the probaility associated to each node, from the 'n_qubit' probability distribution
nodes_probs = compute_nodes_probs(probabilities, basis_states_lists)
# Get the final, computed partition
partition = ['0' if node_prob < 1 / (2*B) else '1' for node_prob in nodes_probs]
return probabilities, nodes_probs, cost, cost_vec, ar_vec, parameters, partition, train_time
n_layers, step_size = 2, 0.98
params_QAOA = 0.1 * np.random.rand(2, n_layers, requires_grad=True); print(f"Initial parameters [QAOA]: {params_QAOA}.")
dev = 'default.qubit'
_, _, _, _, _, _, _, _ = iQAQE_QAOA_Ansatz(graph, n_qubits = n_nodes, n_layers = n_layers, device = dev,
parameters = params_QAOA, B = 4, basis_states_lists = basis_states_lists,
max_iter = 100, step_size = step_size)
Is there anything that jumps out to you as seemingly wrong that could be causing this behaviour [parameters not updating]? Any help would be greatly appreciated!
Thank you in advance,
Afonso
For debugging purposes, here’s the output of qml.about():
Name: PennyLane
Version: 0.34.0
Summary: PennyLane is a Python quantum machine learning library by Xanadu Inc.
Home-page: https://github.com/PennyLaneAI/pennylane
Author:
Author-email:
License: Apache License 2.0
Location: [-]
Requires: appdirs, autograd, autoray, cachetools, networkx, numpy, pennylane-lightning, requests, rustworkx, scipy, semantic-version, toml, typing-extensions
Required-by: PennyLane-Lightning
Platform info: Windows-10-10.0.22621-SP0
Python version: 3.11.5
Numpy version: 1.26.1
Scipy version: 1.11.3
Installed devices:
- default.gaussian (PennyLane-0.34.0)
- default.mixed (PennyLane-0.34.0)
- default.qubit (PennyLane-0.34.0)
- default.qubit.autograd (PennyLane-0.34.0)
- default.qubit.jax (PennyLane-0.34.0)
- default.qubit.legacy (PennyLane-0.34.0)
- default.qubit.tf (PennyLane-0.34.0)
- default.qubit.torch (PennyLane-0.34.0)
- default.qutrit (PennyLane-0.34.0)
- null.qubit (PennyLane-0.34.0)
- lightning.qubit (PennyLane-Lightning-0.34.0)
), so it would make sense that the optimizer doesn’t update the parameters. With a slightly different cost function, this seems to be working properly now! Thank you for bringing this possibility to light!