An introduction to genetic programming with esec demonstrating decision tree induction.

gp_decision_tree.py

from esec import esdl_eval
from esec.species.tgp import Instruction, DecisionInstructionWithState
from esec.landscape.tgp import TGPFitness
 
#define the attributes of a Saturday morning
OUTLOOK = ('SUNNY', 'OVERCAST', 'RAIN')
HUMIDITY = ('HIGH', 'NORMAL')
WINDY = (True, False)
 
#define the state object to store attributes of a Saturday morning
class SaturdayMorning(object):
    def __init__(self, outlook, humidity, windy):
        self.outlook = outlook
        self.humidity = humidity
        self.windy = windy
 
    def classification(self):
        if self.outlook == 'SUNNY':
            return 0 if self.humidity == 'HIGH' else 1
        elif self.outlook == 'OVERCAST':
            return 1
        elif self.outlook == 'RAIN':
            return 0 if self.windy else 1
 
#define the fitness cases
fitness_cases = [SaturdayMorning(outlook, humidity, windy) for outlook in OUTLOOK for humidity in HUMIDITY for windy in WINDY]
 
#define the fitness evaluator for individuals
@esdl_eval
def decision_tree_induction(indiv):
    score = 0
    for fitness_case in fitness_cases:
        if indiv.evaluate(indiv, state=fitness_case, terminals=[0, 1]) == fitness_case.classification():
            score += 1
 
    return TGPFitness([score, len(indiv[0])])
 
#define the instruction set
instructions = [
    # evaluate the 1st, 2nd or 3rd parameter based on the conditions
    DecisionInstructionWithState(lambda fitness_case: 1 if fitness_case.outlook == 'SUNNY' else (2 if fitness_case.outlook == 'OVERCAST' else 3), param_count=3, name='OUT'),
    DecisionInstructionWithState(lambda fitness_case: 1 if fitness_case.humidity == 'HIGH' else 2, param_count=2, name='HUM'),
    DecisionInstructionWithState(lambda fitness_case: 1 if fitness_case.windy else 2, param_count=2, name='WIND'),
]
 
#define the system definition
esdl_system_definition = r'''
    FROM random_tgp(instructions=instructions, terminals=2, deepest=4) SELECT (size) population
    YIELD population
 
    BEGIN generation
        FROM population \
            SELECT (0.9*size) to_cross, (0.02*size) to_mutate, (0.08*size) to_reproduce \
            USING fitness_proportional
 
        FROM to_cross SELECT offspring1 USING crossover_one(deepest_result, terminal_prob=0.1)
        FROM to_mutate SELECT offspring2 USING mutate_random(deepest_result)
 
        FROM offspring1, offspring2, to_reproduce SELECT (size) population
 
        YIELD population
    END generation
    '''
 
#define the experiment configuration
config = {
    'landscape': decision_tree_induction,
    'system': {
        'instructions': instructions,
        'definition': esdl_system_definition,
        'size': 500,
        'deepest_result': 17,
    },
    'monitor': {
        'report': 'gen+births+best+local+best_length+time_delta+best_genome',
        'summary': 'status+best+best_length+best_phenome',
        'limits': {
            'iterations': 50,
            'fitness': TGPFitness([len(fitness_cases), 0]),
        }
    },
}