def experiment_sg(of, maxeval, num_runs, hmax):
results = []
for i in tqdm_notebook(range(num_runs), 'Testing hmax={}'.format(hmax)):
result = ShootAndGo(of, maxeval=maxeval, hmax=hmax).search() # dict with results of one run
result['run'] = i
result['heur'] = 'SG_{}'.format(hmax) # name of the heuristic
result['hmax'] = hmax
results.append(result)
return pd.DataFrame(results, columns=['heur', 'run', 'hmax', 'best_x', 'best_y', 'neval'])
def experiment_random(of, maxeval, num_runs):
results = []
for i in tqdm(range(num_runs), 'Testing maxeval={}'.format(maxeval)):
result = ShootAndGo(of, maxeval, hmax=0).search()
result['run'] = i
result['heur'] = 'Random_{}'.format(maxeval)
result['maxeval'] = maxeval
results.append(result)
return pd.DataFrame(
results,
columns=['heur', 'run', 'maxeval', 'best_x', 'best_y', 'neval'])
def get_overview_heurs(of, maxeval, RD=True, FSA=True, DE=True, GO=True):
overview_heurs = []
if RD:
for hmax in [0, 5, 10, 50, np.inf]:
for RD in [False, True]:
overview_heurs.append(
ShootAndGo(of,
maxeval=maxeval,
hmax=hmax,
random_descent=RD))
if FSA:
for T0 in [1e-10, 1e-2, 1, np.inf]:
for mutation in get_overview_mutation(of):
for n0 in [1, 2, 5]:
overview_heurs.append(
FastSimulatedAnnealing(of,
maxeval=maxeval,
T0=T0,
n0=n0,
alpha=2,
mutation=mutation))
if DE:
for N in [4, 10, 20]:
for CR in [0.2, 0.5, 0.8]:
for F in [0.5, 1, 2]:
overview_heurs.append(
DifferentialEvolution(of,
maxeval=maxeval,
N=N,
CR=CR,
F=F))
if GO:
for mutation in get_overview_mutation(of):
for crossover in [
Crossover(),
UniformMultipoint(1),
RandomCombination()
]:
for N in [1, 2, 5, 10, 30, 100]:
for Tsel1 in [0.5, 1]:
overview_heurs.append(
GeneticOptimization(of,
maxeval,
N=N,
M=N * 3,
Tsel1=Tsel1,
Tsel2=0.1,
mutation=mutation,
crossover=crossover))
return overview_heurs
def experiment(of, maxeval, num_runs, hmax, random_descent):
method = 'RD' if random_descent else 'SD'
results = []
for i in tqdm(range(num_runs),
'Testing method={}, hmax={}'.format(method, hmax)):
result = ShootAndGo(of, maxeval=maxeval,
hmax=hmax).search() # dict with results of one run
result['run'] = i
result['heur'] = 'SG_{}_{}'.format(method,
hmax) # name of the heuristic
result['method'] = method
result['hmax'] = hmax
results.append(result)
return pd.DataFrame(
results,
columns=['heur', 'run', 'method', 'hmax', 'best_x', 'best_y', 'neval'])
# * Shoot & Go heuristic (also known as *Iterated Local Search*, *Random-restart hill climbing*, etc)
# * $hmax \in \{ 0, 1, \ldots, \infty \}$ parameter - maximum number of local searches / hill climbs
# * note that $\mathrm{SG}_{0}$ is pure Random Shooting (Random Search)
#
# * implemented as ``class ShootAndGo(Heuristic)`` in ``src/heur_sg.py``
#
#
# In[12]:
from heur_sg import ShootAndGo
# In[13]:
# Random Shooting for the AirShip initialization...
demo_rs = ShootAndGo(airship, maxeval=100, hmax=0)
# ...and execution:
demo_rs.search()
# # 2. Performance evaluation
#
# ## What is the recommended approach to store and analyze results of your experiments?
#
# 1. Append all relevant statistics from a single run into table (e.g. CSV file in memory or on disk), including all task and heuristic parameters
# 2. Load the table into analytical tool of your choice (**data frame**, Excel or Google Docs spreadsheets, etc.)
# 3. Pivot by relevant parameters, visualize in tables or charts
# ## Demonstration
#
# Neccessary setup first:
# In[7]:
# neighbourhood of x:
N = tsp.get_neighborhood(x, 1)
print(N)
# In[8]:
# decoded neighbours and their objective function values
for xn in N:
print('{} ({}) -> {:.4f}'.format(xn, tsp.decode(xn), tsp.evaluate(xn)))
# **Carefully** mind the difference between encoded solution vector vs decoded city tour and meaning of such neighbourhood.
# ### TSP optimization using Random Shooting ($\mathrm{SG}_{0}$)
# In[9]:
heur = ShootAndGo(tsp, maxeval=1000, hmax=0)
print(heur.search())
# # Assignments:
#
# 1. Find a better performing heuristic (to test TSP implementation on your own).
# 2. Can you improve heuristic performance using any
# 1. **better random point generator**?
# 2. **better neighbourhood generator**?
#
# Use performance measure(s) of your choice (e.g. $FEO$).