Packages

In [1]:
# Python package
import sys
sys.path.append('build')
import numpy as np
import pandas as pd
from sklearn.model_selection import train_test_split, LeaveOneGroupOut
from sklearn.tree import DecisionTreeClassifier
from sklearn import tree
from math import *

# Pysat packages
from pysat.card import *
from pysat.pb import PBEnc
from pysat.formula import WCNF, CNF
from pysat.pb import *
from pysat.examples.rc2 import RC2
from pysat.solvers import Glucose4
from pysat.solvers import Solver, Minisat22

#ours package 
import my_tree as mt
import my_tree

# SAT encoding of "Arenas et al 2022" packages
import utils
import dtree
import gen_dt
import encoder 

# import help function and error function 
from UAI_Functions import W_f, M_s, D_s, Approx_ascendant, Approx_descent, Dichotomie

Learn DT and import the dataset

In [2]:
data = pd.read_csv('tic-tac-toe.csv')
df = data.copy()
ind = df.columns[-1]
y = df[ind]
X = df.drop(ind, axis=1)
X_train, X_test, y_train, y_test = train_test_split(X, y, test_size=0.3)

dt = DecisionTreeClassifier()
dt.fit(X_train, y_train)

my_tree = mt.decision_tree()
my_tree.from_DecisionTreeClassifier(dt)

/home/louenas/anaconda3/lib/python3.9/site-packages/sklearn/utils/deprecation.py:103: FutureWarning: The attribute `n_features_` is deprecated in 1.0 and will be removed in 1.2. Use `n_features_in_` instead.
  warnings.warn(msg, category=FutureWarning)

Example of application of our algorithms and binary search with tic_tac_toe

In [5]:
i = 0# the index of instance
file = 'data_dt'# if you want to change the name, you need to change it also in Dichotomie function
dim = data.shape[1]
ind = X_test.index[i]
pred = dt.predict(X_test)[i]
ins = list(X_test.loc[ind]) + [pred]
instance = X_test.loc[ind]
target = my_tree.take_decision(instance=instance)
dnf = my_tree.to_DNF(target=target, threshold_clauses=False)[0].clauses
phi = my_tree.to_DNF(target=1-target, threshold_clauses=False)[0].clauses
implicant = my_tree.binarized_instance(instance=instance)
n = len(my_tree.bina)
R_d = my_tree.find_direct_reason(instance=instance) # direct reason 
R = my_tree.unbinarized_instance(R_d)
gen_dt.dt_to_file(dimension = dim, dt=dt, filename=file, desired_n_leaves=80, positive_class=pred) 

n_actual_leaves = 77, desired_n_leaves = 80
In [6]:
print('the size of direct reason is: {}'.format(len(R_d))) 

the size of direct reason is: 7
In [7]:
GD = Approx_descent(phi=phi, I_s=R_d, n=n, k=4)
GA = Approx_ascendant(phi=phi, I_s=R_d, n=n, k=4)

print("the output of GD is: {} and GA is: {}".format(GD, GA))

the output of GD is: [-1, -4, -8, -5] and GA is: [-1, -4, -8, -5]

To get the error of GA and GD

In [8]:
print("error of GD is: {} and the error of GA is: {}".format(W_f(phi, GA, n), W_f(phi, GD, n)))

error of GD is: 0.125 and the error of GA is: 0.125

For binary binary search "error"

Note: for more information about the protocol experemental of SAT encoding (see {https://arxiv.org/abs/2207.12213}).

In [9]:
di = Dichotomie(a=0, b=1, k=4, ins=ins, R=R, e=0.001)
print("the error of binary search is: {}".format(1-di))

the error of binary search is: 0.12451171875

Now, if you want test on mnist49

In [10]:
data = pd.read_csv('mnist49.csv')
df = data.copy()
ind = df.columns[-1]
y = df[ind]
X = df.drop(ind, axis=1)
X_train, X_test, y_train, y_test = train_test_split(X, y, test_size=0.3)

dt = DecisionTreeClassifier()
dt.fit(X_train, y_train)

my_tree = mt.decision_tree()
my_tree.from_DecisionTreeClassifier(dt)

/home/louenas/anaconda3/lib/python3.9/site-packages/sklearn/utils/deprecation.py:103: FutureWarning: The attribute `n_features_` is deprecated in 1.0 and will be removed in 1.2. Use `n_features_in_` instead.
  warnings.warn(msg, category=FutureWarning)
In [11]:
i = 2 # the index of instance
file = 'data_dt'
dim = data.shape[1]
ind = X_test.index[i]
pred = dt.predict(X_test)[i]
ins = list(X_test.loc[ind]) + [pred]
instance = X_test.loc[ind]
target = my_tree.take_decision(instance=instance)
dnf = my_tree.to_DNF(target=target, threshold_clauses=False)[0].clauses
phi = my_tree.to_DNF(target=1-target, threshold_clauses=False)[0].clauses
implicant = my_tree.binarized_instance(instance=instance)
n = len(my_tree.bina)
R_d = my_tree.find_direct_reason(instance=instance) # direct reason 
R = my_tree.unbinarized_instance(R_d)
gen_dt.dt_to_file(dimension = dim, dt=dt, filename=file, desired_n_leaves=80, positive_class=pred) 

n_actual_leaves = 292, desired_n_leaves = 80
In [12]:
print('the size of direct reason is: {}'.format(len(R_d))) 

the size of direct reason is: 13
In [13]:
GD = Approx_descent(phi=phi, I_s=R_d, n=n, k=7)
GA = Approx_ascendant(phi=phi, I_s=R_d, n=n, k=7)

print("the output of GD is: {} and GA is: {}".format(GD, GA))

the output of GD is: [1, -41, -13, 87, 71, 17, -11] and GA is: [1, -41, -13, 71, 17, 87, -119]
In [14]:
print("error of GD is: {} and the error of GA is: {}".format(W_f(phi, GA, n), W_f(phi, GD, n)))

error of GD is: 0.2998046875 and the error of GA is: 0.2998046875

Now, if you want to check a binary search

In [ ]:
di = Dichotomie(a=0, b=1, k=3, ins=ins, R=R, e=0.001)
print("the error of binary search is: {}".format(1-di))
In [15]:
print("Sorry, that is not possible")

Sorry, that is not possible