Results and history

Every heuristic returns the same OptimizationResult, so you read and compare runs the same way.

Every heuristic returns an OptimizationResult:

Field	Meaning
best_solution	best feasible solution found
best_value	its objective value
history	one record per iteration; the initial state comes first, so len(history) == iterations + 1
iterations	completed iterations (GA: generations)
elapsed	wall time in seconds
metadata	evaluations (objective calls), termination ('stop'), state (the final State, the one that satisfied stop), plus algorithm extras

metadata['state'] tells you which budget ended the run:

state = sa.metadata['state']
print(sa.metadata['termination'], state.iteration, state.evaluations)
print('final temperature:', sa.metadata['final_temperature'])  # SA extra

The two heuristics record different things in history:

• Simulated Annealing: a float per iteration, the best value so far.
• Genetic Algorithm: a dict per generation:

{'best': float,        # this generation's best value
 'mean': float,        # mean value of the population
 'worst': float,       # worst value of the population
 'solution': genome,   # this generation's best genome
 'best_so_far': float} # best value up to and including this generation

best_so_far gives both heuristics a common convergence measure:

def convergence(result):
    return [r['best_so_far'] if isinstance(r, dict) else r
            for r in result.history]


assert convergence(ga)[-1] == ga.best_value
assert convergence(sa)[-1] == sa.best_value

To plot it (pip install matplotlib):

import matplotlib.pyplot as plt

plt.plot(convergence(sa), label='simulated annealing (per iteration)')
plt.plot(convergence(ga), label='genetic algorithm (per generation)')
plt.xlabel('iteration')
plt.ylabel('best value so far')
plt.legend()
plt.show()

Warning

The two x axes count different things. A GA generation costs population_size evaluations, and an SA iteration costs at most one. To compare the heuristics at equal cost, give both the same max_evaluations stop, as the experiments do.

Recap

• best_solution, best_value, history, iterations, elapsed, metadata.
• best_so_far gives both heuristics a common convergence curve.

Next: reproducibility.