def DistanceVelocity(genotype):
"""Third version of the fitness function. This is the first one used in Kadihasanoglu et al. 2015.
Here we evolve agents based on minimizing their final distance and final velocity."""
final_distances = np.empty(ntrials)
final_velocities = np.empty(ntrials)
jerks = np.empty(ntrials)
i = 0
agent = AgentEnv(genotype, Size, WeightRange, BiasRange, TimeConstMin,
TimeConstMax, InputWeightRange, Dt)
for ts in target_size:
for d in initial_distance:
for v in initial_velocity:
agent.setInitialState(v, d, ts)
agent.Optical_variable = optical_variable
accelerations = []
# While agent hasn't crashed, agent is still moving forward significantly, and not too much time has elapsed
while (agent.Distance > 0) and (agent.Velocity > 0.005) and (
agent.Time < trial_length):
agent.sense()
agent.think()
agent.act()
accelerations.append(agent.Acceleration)
if agent.Distance < 0: # If agent crashed, reset distance to starting position
agent.Distance = d
if agent.Velocity < 0: # If agent finishes moving backwards, reset velocity to starting velocity
agent.Velocity = v
final_distances[i] = agent.Distance / d
final_velocities[i] = agent.Velocity / v
jerks[i] = jerk(accelerations)
i += 1
jweight = 0.
fitness = (
(1 - np.average(final_distances)) +
(1 - np.average(final_velocities))) / 2 - jweight * np.average(jerks)
return fitness
def FinalDistanceNoCrash(genotype):
"""Second version of the fitness function. Here we evolve agents on their ability to end the
trial near the target, WITHOUT crashing. Crashes result in a default fitness of the initial
distance."""
final_distances = np.empty(ntrials)
i = 0
agent = AgentEnv(genotype, Size, WeightRange, BiasRange, TimeConstMin,
TimeConstMax, InputWeightRange, Dt)
for ts in target_size:
for d in initial_distance:
for v in initial_velocity:
agent.setInitialState(v, d, ts)
agent.Optical_variable = optical_variable
# While agent hasn't crashed, agent is still moving forward significantly, and not too much time has elapsed
while (agent.Distance > 0) and (agent.Velocity > 0.005) and (
agent.Time < trial_length):
agent.sense()
agent.think()
agent.act()
if agent.Distance < 0: # if agent crashed, reset distance to starting position
agent.Distance = d
final_distances[
i] = agent.Distance / d # fitness is proportion of distance
i += 1
return (1 - np.average(final_distances))
def FinalDistance(genotype):
"""First version of the fitness function. Here we evolve agents simply on their ability to end
the trial near the target - crashes count, too!"""
final_distances = np.empty(ntrials)
i = 0
agent = AgentEnv(genotype, Size, WeightRange, BiasRange, TimeConstMin,
TimeConstMax, InputWeightRange, Dt)
for ts in target_size:
for d in initial_distance:
for v in initial_velocity:
agent.setInitialState(v, d, ts)
agent.Optical_variable = optical_variable
# While distance is still positive, agent is still moving forward significantly, and not too much time has elapsed
while (agent.Distance > 0) and (agent.Velocity > 0.005) and (
agent.Time < trial_length):
agent.sense()
agent.think()
agent.act()
final_distances[i] = agent.Distance
i += 1
return (1 - np.average(final_distances))
save(filename, mga)
print('%f sec elapsed so far \n' % (time.time() - start))
# =========================================== RUNTIME =========================================
#run_continue('DistanceVelocityJerk_V2_P150_T168_G175_2021-04-28', 25, 25)
# Set up
start = time.time()
print('Number of Evaluation Trials: %i' % ntrials)
# Run simulation
mga = Microbial(fitnessFunction, Population, GenotypeLength, RecombProb,
MutatProb)
mga.runTournaments(Tournaments)
# Save data
ffname = str(mga.fitnessFunction.__name__)
generation = int(mga.generationsRun)
filename = '%s_V%i_test' % (ffname, optical_variable)
save(filename, mga)
#Report runtime
print('TOTAL TIME ELAPSED: %i sec' % (int(time.time() - start)))
# Show graphs and save them too
mga.showFitness()
mga.showDiversity()
# Show trajectories of best individual
agent = AgentEnv(mga.bestIndividual, Size, WeightRange, BiasRange,
TimeConstMin, TimeConstMax, InputWeightRange, Dt)
agent.showTrajectory(optical_variable, target_size[0], initial_distance[0],
initial_velocity[0])
def trialAnalysis(self, traj_params=(), show=True):
"""Suggested value for jweight are 0 or 1000 (KBB15). traj_params should be a
tuple of (initial velocity, initial distance, target size)."""
# Step 1: Get genotype of best individual
genotype = self.bestIndividual
# Step 2: Run simulation
final_distances = np.empty(ntrials)
final_velocities = np.empty(ntrials)
jerks = np.empty(ntrials)
trials = 0
successes = 0
crashes = 0
early_stops = 0
timeouts = 0
agent = AgentEnv(genotype, Size, WeightRange, BiasRange, TimeConstMin, TimeConstMax, InputWeightRange, Dt)
for ts in target_size:
for d in initial_distance:
for v in initial_velocity:
agent.setInitialState(v, d, ts)
agent.Optical_variable = self.Optical_variable
acceleration = []
distance = []
velocity = []
optical = []
# Simulation loop
i = 0
while (agent.Distance > 0) and (agent.Velocity > 0.005) and (agent.Time < trial_length):
agent.sense()
agent.think()
agent.act()
i += 1
acceleration.append(agent.Acceleration) # Record data
distance.append(agent.Distance)
velocity.append(agent.Velocity)
optical.append(agent.Optical_info[agent.Optical_variable])
if show == True and (v, d, ts) == traj_params: # If this is the trial we want to visualize, record data for plotting
series = [distance, velocity, acceleration, optical]
time = agent.Time
if agent.Velocity < 0: # If agent finishes moving backwards, reset velocity to starting velocity
agent.Velocity = v
if agent.Distance > 15: # If agent stopped prematurely
early_stops += 1
elif agent.Distance < 0: # If agent crashed, reset distance to starting position
agent.Distance = d
crashes += 1
elif agent.Time > trial_length+Dt: # If agent never stops/times out
timeouts += 1
else:
successes += 1
final_distances[trials] = agent.Distance/d # Record performance data for fitness function
final_velocities[trials] = agent.Velocity/v
jerks[trials] = jerk(acceleration)
trials += 1
# Plot data
if show == True:
for i in range(len(series)): # This is necessary because of an inconsistent error I was getting
time = np.arange(0, trial_length, agent.Dt)
l = len(series[i])
time = time[:l]
plt.plot(time, series[i])
ov_labels = ['Image size', 'Image expansion rate', 'Tau', 'Tau-dot', 'Proportional Rate']
labels = ['Distance (m)', 'Velocity (m/sec)', 'Acceleration (m/sec^2)]', ov_labels[agent.Optical_variable] ]
plt.xlabel('Time (sec)')
plt.ylabel(labels[i])
plt.title('%s-Guided Agent' % (ov_labels[agent.Optical_variable]))
plt.show()
self.successes = successes
self.crashes = crashes
self.earlyStops = early_stops
self.timeouts = timeouts
def perturbationAnalysis(self, ptype, perturbations, traj_params=(), show=True):
"""Valid value for ptype are 'Position', 'Velocity', 'Mapping', and 'Delay'. Perturbations
should be a vector of perturbation values to be iterated through and applied at every step. The
length should be trial_length/Dt. The exception for this is in the case of Delay perturbations,
in which case perturbations should be an integer representing the number of time steps to delay
motor output by. Suggested value for jweight are 0 or 1000 (KBB15). traj_params should be a
tuple of (initial velocity, initial distance, target size)."""
# Step 1: Get genotype of best individual
genotype = self.bestIndividual
# Step 2: Run simulation
final_distances = np.empty(ntrials)
final_velocities = np.empty(ntrials)
jerks = np.empty(ntrials)
trials = 0
crashes = 0
early_stops = 0
timeouts = 0
agent = AgentEnv(genotype, Size, WeightRange, BiasRange, TimeConstMin, TimeConstMax, InputWeightRange, Dt)
for ts in target_size:
for d in initial_distance:
for v in initial_velocity:
agent.setInitialState(v, d, ts)
agent.Optical_variable = self.Optical_variable
acceleration = []
distance = []
velocity = []
optical = []
if ptype == 'Delay':
outputs = []
for o in range(perturbations):
outputs.append(0)
# Simulation loop
i = 0
while (agent.Distance > 0) and (agent.Velocity > 0.005) and (agent.Time < trial_length):
agent.sense()
agent.think()
if ptype == 'Delay':
outputs.append(agent.output) # Take self.output and append it to end of the list
agent.output = outputs.pop(0) # Remove first element from list and set it to self.output
agent.act()
if ptype == 'Position':
agent.Distance += perturbations[i]
if ptype == 'Velocity':
agent.Velocity += perturbations[i]
if ptype == 'Mapping':
agent.brake_effectiveness = perturbations[i]
i += 1
acceleration.append(agent.Acceleration) # Record data
distance.append(agent.Distance)
velocity.append(agent.Velocity)
optical.append(agent.Optical_info[agent.Optical_variable])
if show == True and (v, d, ts) == traj_params: # If this is the trial we want to visualize, record data for plotting
series = [distance, velocity, acceleration, optical]
time = agent.Time
if agent.Distance < 0: # If agent crashed, reset distance to starting position
agent.Distance = d
crashes += 1
if agent.Distance > 15: # If agent stopped prematurely
early_stops += 1
if agent.Time > trial_length: # If agent never stops/times out
timeouts += 1
if agent.Velocity < 0: # If agent finishes moving backwards, reset velocity to starting velocity
agent.Velocity = v
final_distances[trials] = agent.Distance/d # Record performance data for fitness function
final_velocities[trials] = agent.Velocity/v
jerks[trials] = jerk(acceleration)
trials += 1
# Plot data
fitness= ( (1-np.average(final_distances)) + (1-np.average(final_velocities)) )/2 - 1000*np.average(jerks)
if show == True:
for i in range(len(series)): # This is necessary because of an inconsistent error I was getting
time = np.arange(0, trial_length, agent.Dt)
l = len(series[i])
time = time[:l]
plt.plot(time, series[i])
ov_labels = ['Image size', 'Image expansion rate', 'Tau', 'Tau-dot', 'Proportional Rate']
labels = ['Distance (m)', 'Velocity (m/sec)', 'Acceleration (m/sec^2)]', ov_labels[agent.Optical_variable] ]
plt.xlabel('Time (sec)')
plt.ylabel(labels[i])
plt.title('%s Perturbation (%s agent)' % (ptype, ov_labels[agent.Optical_variable]))
plt.show()
print('Original Fitness: %f' % self.bestFitness)
print('Perturbed Fitness: %f' % fitness)
successes = trials - crashes - early_stops - timeouts
if ptype=='Delay':
self.delayPerturbStats = fitness, successes, crashes, early_stops, timeouts
self.delayPerturbation = perturbations
elif ptype=='Position':
self.positionPerturbStats = fitness, successes, crashes, early_stops, timeouts
self.positionPerturbation = perturbations
elif ptype=='Velocity':
self.velocityPerturbStats = fitness, successes, crashes, early_stops, timeouts
self.velocityPerturbation = perturbations
elif ptype=='Mapping':
self.mappingPerturbStats = fitness, successes, crashes, early_stops, timeouts
self.mappingPerturbation = perturbations
else:
print('Ptype argument not recognized!')