qtt_max.py

"""The demo of using ttopt for maximization. Example with QTT.

We'll find the maximum for the 10-dimensional Alpine function with vectorized
input. The target function for maximization has the form f(X), where input X is
the [samples, dimension] numpy array.

Run it from the root of the project as "python demo/qtt_max.py".

As a result of the script work we expect the output in console like this:
"
...
Alpine-10d | evals=1.00e+05 | t_cur=1.65e-01 | y= 8.715206e+01
----------------------------------------------------------------------
Alpine-10d | evals=1.00e+05 | t_all=2.22e+00 | y= 8.715206e+01 
"

"""
import numpy as np


from ttopt import TTOpt
from ttopt import ttopt_init


np.random.seed(42)


d = 10                      # Number of function dimensions:
rank = 4                    # Maximum TT-rank while cross-like iterations
def f(X):                   # Target function
    return np.sum(np.abs(X * np.sin(X) + 0.1 * X), axis=1)


# We initialize the TTOpt class instance with the correct parameters:
tto = TTOpt(
    f=f,                    # Function for maximization. X is [samples, dim]
    d=d,                    # Number of function dimensions
    a=-10.,                 # Grid lower bound (number or list of len d)
    b=+10.,                 # Grid upper bound (number or list of len d)
    p=2,                    # The grid size factor (there will n=p^q points)
    q=20,                   # The grid size factor (there will n=p^q points)
    evals=1.E+5,            # Number of function evaluations
    name='Alpine',          # Function name for log (this is optional)
    with_log=True)


# And now we launching the maximizer:
tto.optimize(rank, is_max=True)


# We can extract the results of the computation:
x = tto.x_opt          # The found value of the maximum of the function (x)
y = tto.y_opt          # The found value of the maximum of the function (y=f(x))
k_c = tto.k_cache      # Total number of cache usage (should be 0 in this demo)
k_e = tto.k_evals      # Total number of requests to func (is always = evals)
k_t = tto.k_total      # Total number of requests (k_cache + k_evals)
t_f = tto.t_evals_mean # Average time spent to real function call for 1 point
                       # ... (see "ttopt.py" and docs for more details)


# We log the final state:
print('-' * 70 + '\n' + tto.info() +'\n\n')