def test_lvoc_features_function(self):
m1 = pnl.TransferMechanism(
input_states=["InputState A", "InputState B"])
m2 = pnl.TransferMechanism()
c = pnl.Composition()
c.add_node(m1, required_roles=pnl.NodeRole.INPUT)
c.add_node(m2, required_roles=pnl.NodeRole.INPUT)
c._analyze_graph()
lvoc = pnl.OptimizationControlMechanism(
agent_rep=pnl.RegressionCFA,
features=[
m1.input_states[0], m1.input_states[1], m2.input_state, m2
],
feature_function=pnl.LinearCombination(offset=10.0),
objective_mechanism=pnl.ObjectiveMechanism(monitor=[m1, m2]),
function=pnl.GradientOptimization(max_iterations=1),
control_signals=[(pnl.SLOPE, m1), (pnl.SLOPE, m2)])
c.add_node(lvoc)
input_dict = {m1: [[1], [1]], m2: [1]}
c.run(inputs=input_dict)
assert len(lvoc.input_states) == 5
for i in range(1, 5):
assert lvoc.input_states[i].function.offset == 10.0
def test_model_based_ocm_no_simulations(self):
A = pnl.ProcessingMechanism(name='A')
B = pnl.ProcessingMechanism(name='B',
function=pnl.SimpleIntegrator(rate=1))
comp = pnl.Composition(name='comp')
comp.add_linear_processing_pathway([A, B])
control_signal = pnl.ControlSignal(
projections=[(pnl.SLOPE, A)],
function=pnl.Linear,
variable=1.0,
allocation_samples=[1, 2, 3],
intensity_cost_function=pnl.Linear(slope=0.))
objective_mech = pnl.ObjectiveMechanism(monitor=[B])
ocm = pnl.OptimizationControlMechanism(
agent_rep=comp,
features=[A.input_state],
objective_mechanism=objective_mech,
function=pnl.GridSearch(),
num_estimates=1,
control_signals=[control_signal],
search_statefulness=False,
)
comp.add_controller(ocm)
inputs = {A: [[[1.0]]]}
comp.run(inputs=inputs, num_trials=1)
# initial 1 + each allocation sample (1, 2, 3) integrated
assert B.parameters.value.get(comp) == 7
def test_model_based_ocm_with_buffer(self):
A = pnl.ProcessingMechanism(name='A')
B = pnl.ProcessingMechanism(name='B')
comp = pnl.Composition(name='comp', controller_mode=pnl.BEFORE)
comp.add_linear_processing_pathway([A, B])
search_range = pnl.SampleSpec(start=0.25, stop=0.75, step=0.25)
control_signal = pnl.ControlSignal(
projections=[(pnl.SLOPE, A)],
function=pnl.Linear,
variable=1.0,
allocation_samples=search_range,
intensity_cost_function=pnl.Linear(slope=0.))
objective_mech = pnl.ObjectiveMechanism(monitor=[B])
ocm = pnl.OptimizationControlMechanism(
agent_rep=comp,
features=[A.input_state],
feature_function=pnl.Buffer(history=2),
objective_mechanism=objective_mech,
function=pnl.GridSearch(),
control_signals=[control_signal])
objective_mech.log.set_log_conditions(pnl.OUTCOME)
comp.add_controller(ocm)
inputs = {A: [[[1.0]], [[2.0]], [[3.0]]]}
for i in range(1, len(ocm.input_states)):
ocm.input_states[i].function.reinitialize()
comp.run(inputs=inputs, retain_old_simulation_data=True)
log = objective_mech.log.nparray_dictionary()
# "outer" composition
assert np.allclose(log["comp"][pnl.OUTCOME], [[0.75], [1.5], [2.25]])
# preprocess to ignore control allocations
log_parsed = {}
for key, value in log.items():
cleaned_key = re.sub(r'comp-sim-(\d).*', r'\1', key)
log_parsed[cleaned_key] = value
# First round of simulations is only one trial.
# (Even though the feature fn is a Buffer, there is no history yet)
for i in range(0, 3):
assert len(log_parsed[str(i)]["Trial"]) == 1
# Second and third rounds of simulations are two trials.
# (The buffer has history = 2)
for i in range(3, 9):
assert len(log_parsed[str(i)]["Trial"]) == 2
def test_lvoc(self):
m1 = pnl.TransferMechanism(
input_states=["InputState A", "InputState B"])
m2 = pnl.TransferMechanism()
c = pnl.Composition()
c.add_node(m1, required_roles=pnl.NodeRole.INPUT)
c.add_node(m2, required_roles=pnl.NodeRole.INPUT)
c._analyze_graph()
lvoc = pnl.OptimizationControlMechanism(
agent_rep=pnl.RegressionCFA,
features=[m1.input_states[0], m1.input_states[1], m2.input_state],
objective_mechanism=pnl.ObjectiveMechanism(monitor=[m1, m2]),
function=pnl.GridSearch(max_iterations=1),
control_signals=[(pnl.SLOPE, m1), (pnl.SLOPE, m2)])
c.add_node(lvoc)
input_dict = {m1: [[1], [1]], m2: [1]}
c.run(inputs=input_dict)
assert len(lvoc.input_states) == 4
def test_grid_search_random_selection(self):
A = pnl.ProcessingMechanism(name='A')
A.log.set_log_conditions(items="mod_slope")
B = pnl.ProcessingMechanism(name='B', function=pnl.Logistic())
comp = pnl.Composition(name='comp')
comp.add_linear_processing_pathway([A, B])
search_range = pnl.SampleSpec(start=15., stop=35., step=5)
control_signal = pnl.ControlSignal(
projections=[(pnl.SLOPE, A)],
function=pnl.Linear,
variable=1.0,
allocation_samples=search_range,
intensity_cost_function=pnl.Linear(slope=0.))
objective_mech = pnl.ObjectiveMechanism(monitor=[B])
ocm = pnl.OptimizationControlMechanism(
agent_rep=comp,
features=[A.input_state],
objective_mechanism=objective_mech,
function=pnl.GridSearch(select_randomly_from_optimal_values=True),
control_signals=[control_signal])
comp.add_controller(ocm)
inputs = {A: [[[1.0]]]}
comp.run(inputs=inputs, num_trials=10, execution_id='outer_comp')
log_arr = A.log.nparray_dictionary()
# control signal value (mod slope) is chosen randomly from all of the control signal values
# that correspond to a net outcome of 1
assert np.allclose([[1.], [15.], [15.], [20.], [20.], [15.], [20.],
[25.], [15.], [35.]],
log_arr['outer_comp']['mod_slope'])
def test_model_based_num_estimates(self):
A = pnl.ProcessingMechanism(name='A')
B = pnl.ProcessingMechanism(name='B',
function=pnl.SimpleIntegrator(rate=1))
comp = pnl.Composition(name='comp')
comp.add_linear_processing_pathway([A, B])
search_range = pnl.SampleSpec(start=0.25, stop=0.75, step=0.25)
control_signal = pnl.ControlSignal(
projections=[(pnl.SLOPE, A)],
function=pnl.Linear,
variable=1.0,
allocation_samples=search_range,
intensity_cost_function=pnl.Linear(slope=0.))
objective_mech = pnl.ObjectiveMechanism(monitor=[B])
ocm = pnl.OptimizationControlMechanism(
agent_rep=comp,
features=[A.input_state],
objective_mechanism=objective_mech,
function=pnl.GridSearch(),
num_estimates=5,
control_signals=[control_signal])
comp.add_controller(ocm)
inputs = {A: [[[1.0]]]}
comp.run(inputs=inputs, num_trials=2)
assert np.allclose(
comp.simulation_results,
[[np.array([2.25])], [np.array([3.5])], [np.array([4.75])],
[np.array([3.])], [np.array([4.25])], [np.array([5.5])]])
assert np.allclose(comp.results,
[[np.array([1.])], [np.array([1.75])]])
def test_model_based_ocm_before(self, benchmark, mode):
A = pnl.ProcessingMechanism(name='A')
B = pnl.ProcessingMechanism(name='B')
comp = pnl.Composition(name='comp', controller_mode=pnl.BEFORE)
comp.add_linear_processing_pathway([A, B])
search_range = pnl.SampleSpec(start=0.25, stop=0.75, step=0.25)
control_signal = pnl.ControlSignal(
projections=[(pnl.SLOPE, A)],
function=pnl.Linear,
variable=1.0,
allocation_samples=search_range,
intensity_cost_function=pnl.Linear(slope=0.))
objective_mech = pnl.ObjectiveMechanism(monitor=[B])
ocm = pnl.OptimizationControlMechanism(
agent_rep=comp,
features=[A.input_state],
objective_mechanism=objective_mech,
function=pnl.GridSearch(),
control_signals=[control_signal])
# objective_mech.log.set_log_conditions(pnl.OUTCOME)
comp.add_controller(ocm)
inputs = {A: [[[1.0]], [[2.0]], [[3.0]]]}
comp.run(inputs=inputs, bin_execute=mode)
# objective_mech.log.print_entries(pnl.OUTCOME)
assert np.allclose(
comp.results,
[[np.array([0.75])], [np.array([1.5])], [np.array([2.25])]])
benchmark(comp.run, inputs, bin_execute=mode)
output_states=[
pnl.DDM_OUTPUT.PROBABILITY_UPPER_THRESHOLD,
pnl.DDM_OUTPUT.PROBABILITY_LOWER_THRESHOLD
])
task_decision.set_log_conditions('func_drift_rate')
task_decision.set_log_conditions('mod_drift_rate')
task_decision.set_log_conditions('PROBABILITY_LOWER_THRESHOLD')
task_decision.set_log_conditions('PROBABILITY_UPPER_THRESHOLD')
print("task decision loggables: ", task_decision.loggable_items)
lvoc = pnl.LVOCControlMechanism(
name='LVOC ControlMechanism',
feature_predictors={pnl.SHADOW_EXTERNAL_INPUTS: [color_stim, word_stim]},
objective_mechanism=pnl.ObjectiveMechanism(
name='LVOC ObjectiveMechanism',
monitored_output_states=[task_decision, reward],
function=objective_function),
prediction_terms=[pnl.PV.FC, pnl.PV.COST],
terminal_objective_mechanism=True,
# learning_function=pnl.BayesGLM(mu_0=0, sigma_0=0.1),
learning_function=pnl.BayesGLM(mu_0=-0.17, sigma_0=0.11),
# function=pnl.GradientOptimization(
# convergence_criterion=pnl.VALUE,
# convergence_threshold=0.001,
# step_size=1,
# annealing_function= lambda x,y : x / np.sqrt(y),
# # direction=pnl.ASCENT
# ),
function=pnl.GridSearch,
)
DECISION = pnl.DDM(
name="DECISION",
function=pnl.DriftDiffusionAnalytical(default_variable=[[0.0]]),
input_ports=[{
pnl.NAME:
pnl.ARRAY,
pnl.VARIABLE: [[0.0, 0.0]],
pnl.FUNCTION:
pnl.Reduce(default_variable=[[0.0, 0.0]], weights=[1, -1]),
}],
)
Conflict_Monitor = pnl.ObjectiveMechanism(
name="Conflict_Monitor",
function=pnl.Energy(matrix=[[0, -2.5], [-2.5, 0]],
default_variable=[[0.0, 0.0]]),
monitor=[OUTPUT],
default_variable=[[0.0, 0.0]],
)
CONTROL = pnl.ControlMechanism(
name="CONTROL",
default_allocation=[0.5],
function=pnl.DefaultAllocationFunction(default_variable=[[1.0]]),
monitor_for_control=[],
objective_mechanism=Conflict_Monitor,
control=[{
pnl.NAME: pnl.GAIN,
pnl.MECHANISM: TASK
}],
)
def get_stroop_model(unit_noise_std=.01, dec_noise_std=.1):
# model params
integration_rate = 1
hidden_func = pnl.Logistic(gain=1.0, x_0=4.0)
# input layer, color and word
reward = pnl.TransferMechanism(name='reward')
punish = pnl.TransferMechanism(name='punish')
inp_clr = pnl.TransferMechanism(
size=N_UNITS, function=pnl.Linear, name='COLOR INPUT'
)
inp_wrd = pnl.TransferMechanism(
size=N_UNITS, function=pnl.Linear, name='WORD INPUT'
)
# task layer, represent the task instruction; color naming / word reading
inp_task = pnl.TransferMechanism(
size=N_UNITS, function=pnl.Linear, name='TASK'
)
# hidden layer for color and word
hid_clr = pnl.TransferMechanism(
size=N_UNITS,
function=hidden_func,
integrator_mode=True,
integration_rate=integration_rate,
# noise=pnl.NormalDist(standard_deviation=unit_noise_std).function,
noise=pnl.NormalDist(standard_deviation=unit_noise_std),
name='COLORS HIDDEN'
)
hid_wrd = pnl.TransferMechanism(
size=N_UNITS,
function=hidden_func,
integrator_mode=True,
integration_rate=integration_rate,
# noise=pnl.NormalDist(standard_deviation=unit_noise_std).function,
noise=pnl.NormalDist(standard_deviation=unit_noise_std),
name='WORDS HIDDEN'
)
# output layer
output = pnl.TransferMechanism(
size=N_UNITS,
function=pnl.Logistic,
integrator_mode=True,
integration_rate=integration_rate,
# noise=pnl.NormalDist(standard_deviation=unit_noise_std).function,
noise=pnl.NormalDist(standard_deviation=unit_noise_std),
name='OUTPUT'
)
# decision layer, some accumulator
signalSearchRange = pnl.SampleSpec(start=0.05, stop=5, step=0.05)
decision = pnl.DDM(name='Decision',
input_format=pnl.ARRAY,
function=pnl.DriftDiffusionAnalytical(drift_rate=1,
threshold =1,
noise=1,
starting_point=0,
t0=0.35),
output_ports=[pnl.RESPONSE_TIME,
pnl.PROBABILITY_UPPER_THRESHOLD,
pnl.PROBABILITY_LOWER_THRESHOLD]
)
driftrate_control_signal = pnl.ControlSignal(projections=[(pnl.SLOPE, inp_clr)],
variable=1.0,
intensity_cost_function=pnl.Exponential(rate=1),#pnl.Exponential(rate=0.8),#pnl.Exponential(rate=1),
allocation_samples=signalSearchRange)
threshold_control_signal = pnl.ControlSignal(projections=[(pnl.THRESHOLD, decision)],
variable=1.0,
intensity_cost_function=pnl.Linear(slope=0),
allocation_samples=signalSearchRange)
reward_rate = pnl.ObjectiveMechanism(function=pnl.LinearCombination(operation=pnl.PRODUCT,
exponents=[[1],[1],[-1]]),
monitor=[reward,
decision.output_ports[pnl.PROBABILITY_UPPER_THRESHOLD],
decision.output_ports[pnl.RESPONSE_TIME]])
punish_rate = pnl.ObjectiveMechanism(function=pnl.LinearCombination(operation=pnl.PRODUCT,
exponents=[[1],[1],[-1]]),
monitor=[punish,
decision.output_ports[pnl.PROBABILITY_LOWER_THRESHOLD],
decision.output_ports[pnl.RESPONSE_TIME]])
objective_mech = pnl.ObjectiveMechanism(function=pnl.LinearCombination(operation=pnl.SUM,
weights=[[1],[-1]]),
monitor=[reward_rate, punish_rate])
# objective_mech = pnl.ObjectiveMechanism(function=object_function,
# monitor=[reward,
# punish,
# decision.output_ports[pnl.PROBABILITY_UPPER_THRESHOLD],
# decision.output_ports[pnl.PROBABILITY_LOWER_THRESHOLD],
# (decision.output_ports[pnl.RESPONSE_TIME])])
# PROJECTIONS, weights copied from cohen et al (1990)
wts_clr_ih = pnl.MappingProjection(
matrix=[[2.2, -2.2], [-2.2, 2.2]], name='COLOR INPUT TO HIDDEN')
wts_wrd_ih = pnl.MappingProjection(
matrix=[[2.6, -2.6], [-2.6, 2.6]], name='WORD INPUT TO HIDDEN')
wts_clr_ho = pnl.MappingProjection(
matrix=[[1.3, -1.3], [-1.3, 1.3]], name='COLOR HIDDEN TO OUTPUT')
wts_wrd_ho = pnl.MappingProjection(
matrix=[[2.5, -2.5], [-2.5, 2.5]], name='WORD HIDDEN TO OUTPUT')
wts_tc = pnl.MappingProjection(
matrix=[[4.0, 4.0], [0, 0]], name='COLOR NAMING')
wts_tw = pnl.MappingProjection(
matrix=[[0, 0], [4.0, 4.0]], name='WORD READING')
# build the model
model = pnl.Composition(name='STROOP model')
model.add_node(decision, required_roles=pnl.NodeRole.OUTPUT)
model.add_node(reward, required_roles=pnl.NodeRole.OUTPUT)
model.add_node(punish, required_roles=pnl.NodeRole.OUTPUT)
model.add_linear_processing_pathway([inp_clr, wts_clr_ih, hid_clr])
model.add_linear_processing_pathway([inp_wrd, wts_wrd_ih, hid_wrd])
model.add_linear_processing_pathway([hid_clr, wts_clr_ho, output])
model.add_linear_processing_pathway([hid_wrd, wts_wrd_ho, output])
model.add_linear_processing_pathway([inp_task, wts_tc, hid_clr])
model.add_linear_processing_pathway([inp_task, wts_tw, hid_wrd])
model.add_linear_processing_pathway([output, pnl.IDENTITY_MATRIX, decision]) # 3/15/20
# model.add_linear_processing_pathway([output, [[1,-1]], (decision, pnl.NodeRole.OUTPUT)]) # 3/15/20
# model.add_linear_processing_pathway([output, [[1],[-1]], decision]) # 3/15/20
model.add_nodes([reward_rate, punish_rate])
controller = pnl.OptimizationControlMechanism(agent_rep=model,
features=[inp_clr.input_port,
inp_wrd.input_port,
inp_task.input_port,
reward.input_port,
punish.input_port],
feature_function=pnl.AdaptiveIntegrator(rate=0.1),
objective_mechanism=objective_mech,
function=pnl.GridSearch(),
control_signals=[driftrate_control_signal,
threshold_control_signal])
model.add_controller(controller=controller)
# collect the node handles
nodes = [inp_clr, inp_wrd, inp_task, hid_clr, hid_wrd, output, decision, reward, punish,controller]
metadata = [integration_rate, dec_noise_std, unit_noise_std]
return model, nodes, metadata
c_v_FitzHughNagumo=1.0,
d_v_FitzHughNagumo=0.0,
e_v_FitzHughNagumo=-1.0,
f_v_FitzHughNagumo=1.0,
a_w_FitzHughNagumo=1.0,
b_w_FitzHughNagumo=-1.0,
c_w_FitzHughNagumo=0.0,
t_0_FitzHughNagumo=0.0,
base_level_gain=
G, # Additionally, we set the parameters k and G to compute the gain equation
scaling_factor_gain=k,
initial_v_FitzHughNagumo=initial_v, # Initialize v
initial_w_FitzHughNagumo=initial_w, # Initialize w
objective_mechanism=pnl.ObjectiveMechanism(
function=pnl.Linear,
monitor=[(
decision_layer, # Project output of T1 and T2 but not distractor from decision layer to LC
np.array([[lcwt], [lcwt], [0.0]]))],
name='Combine values'),
modulated_mechanisms=[decision_layer, response_layer
], # Modulate gain of decision & response layers
name='LC')
# Log value of LC
LC.set_log_conditions('value')
# Set initial gain to G + k*initial_w, when the System runs the very first time,
# since the decison layer executes before the LC and hence needs one initial gain value to start with.
for output_port in LC.output_ports:
output_port.value *= G + k * initial_w
task = pnl.Composition()
def test_evc(self):
# Mechanisms
Input = pnl.TransferMechanism(name='Input')
reward = pnl.TransferMechanism(
output_states=[pnl.RESULT, pnl.OUTPUT_MEAN, pnl.OUTPUT_VARIANCE],
name='reward')
Decision = pnl.DDM(function=pnl.DriftDiffusionAnalytical(
drift_rate=(1.0,
pnl.ControlProjection(function=pnl.Linear,
control_signal_params={
pnl.ALLOCATION_SAMPLES:
np.arange(0.1, 1.01, 0.3)
})),
threshold=(1.0,
pnl.ControlProjection(function=pnl.Linear,
control_signal_params={
pnl.ALLOCATION_SAMPLES:
np.arange(0.1, 1.01, 0.3)
})),
noise=0.5,
starting_point=0,
t0=0.45),
output_states=[
pnl.DECISION_VARIABLE, pnl.RESPONSE_TIME,
pnl.PROBABILITY_UPPER_THRESHOLD
],
name='Decision')
comp = pnl.Composition(name="evc")
comp.add_node(reward, required_roles=[pnl.NodeRole.OUTPUT])
comp.add_node(Decision, required_roles=[pnl.NodeRole.OUTPUT])
task_execution_pathway = [Input, pnl.IDENTITY_MATRIX, Decision]
comp.add_linear_processing_pathway(task_execution_pathway)
comp.add_controller(controller=pnl.OptimizationControlMechanism(
agent_rep=comp,
features=[Input.input_state, reward.input_state],
feature_function=pnl.AdaptiveIntegrator(rate=0.5),
objective_mechanism=pnl.ObjectiveMechanism(
function=pnl.LinearCombination(operation=pnl.PRODUCT),
monitor=[
reward, Decision.output_states[
pnl.PROBABILITY_UPPER_THRESHOLD],
(Decision.output_states[pnl.RESPONSE_TIME], -1, 1)
]),
function=pnl.GridSearch(),
control_signals=[("drift_rate", Decision), ("threshold",
Decision)]))
comp.enable_controller = True
comp._analyze_graph()
stim_list_dict = {Input: [0.5, 0.123], reward: [20, 20]}
comp.run(inputs=stim_list_dict, retain_old_simulation_data=True)
# Note: Removed decision variable OutputState from simulation results because sign is chosen randomly
expected_sim_results_array = [[[10.], [10.0], [0.0], [0.48999867],
[0.50499983]],
[[10.], [10.0], [0.0], [1.08965888],
[0.51998934]],
[[10.], [10.0], [0.0], [2.40680493],
[0.53494295]],
[[10.], [10.0], [0.0], [4.43671978],
[0.549834]],
[[10.], [10.0], [0.0], [0.48997868],
[0.51998934]],
[[10.], [10.0], [0.0], [1.08459402],
[0.57932425]],
[[10.], [10.0], [0.0], [2.36033556],
[0.63645254]],
[[10.], [10.0], [0.0], [4.24948962],
[0.68997448]],
[[10.], [10.0], [0.0], [0.48993479],
[0.53494295]],
[[10.], [10.0], [0.0], [1.07378304],
[0.63645254]],
[[10.], [10.0], [0.0], [2.26686573],
[0.72710822]],
[[10.], [10.0], [0.0], [3.90353015],
[0.80218389]],
[[10.], [10.0], [0.0], [0.4898672],
[0.549834]],
[[10.], [10.0], [0.0], [1.05791834],
[0.68997448]],
[[10.], [10.0], [0.0], [2.14222978],
[0.80218389]],
[[10.], [10.0], [0.0], [3.49637662],
[0.88079708]],
[[15.], [15.0], [0.0], [0.48999926],
[0.50372993]],
[[15.], [15.0], [0.0], [1.08981011],
[0.51491557]],
[[15.], [15.0], [0.0], [2.40822035],
[0.52608629]],
[[15.], [15.0], [0.0], [4.44259627],
[0.53723096]],
[[15.], [15.0], [0.0], [0.48998813],
[0.51491557]],
[[15.], [15.0], [0.0], [1.0869779],
[0.55939819]],
[[15.], [15.0], [0.0], [2.38198336],
[0.60294711]],
[[15.], [15.0], [0.0], [4.33535807],
[0.64492386]],
[[15.], [15.0], [0.0], [0.48996368],
[0.52608629]],
[[15.], [15.0], [0.0], [1.08085171],
[0.60294711]],
[[15.], [15.0], [0.0], [2.32712843],
[0.67504223]],
[[15.], [15.0], [0.0], [4.1221271],
[0.7396981]],
[[15.], [15.0], [0.0], [0.48992596],
[0.53723096]],
[[15.], [15.0], [0.0], [1.07165729],
[0.64492386]],
[[15.], [15.0], [0.0], [2.24934228],
[0.7396981]],
[[15.], [15.0], [0.0], [3.84279648],
[0.81637827]]]
for simulation in range(len(expected_sim_results_array)):
assert np.allclose(
expected_sim_results_array[simulation],
# Note: Skip decision variable OutputState
comp.simulation_results[simulation][0:3] +
comp.simulation_results[simulation][4:6])
expected_results_array = [[[20.0], [20.0], [0.0], [1.0],
[2.378055160151634], [0.9820137900379085]],
[[20.0], [20.0], [0.0], [0.1],
[0.48999967725112503],
[0.5024599801509442]]]
for trial in range(len(expected_results_array)):
np.testing.assert_allclose(
comp.results[trial],
expected_results_array[trial],
atol=1e-08,
err_msg='Failed on expected_output[{0}]'.format(trial))
c.add_node(color_task, required_roles=pnl.NodeRole.ORIGIN)
c.add_node(word_task, required_roles=pnl.NodeRole.ORIGIN)
c.add_node(reward)
c.add_node(task_decision)
c.add_projection(sender=color_task, receiver=task_decision)
c.add_projection(sender=word_task, receiver=task_decision)
lvoc = pnl.OptimizationControlMechanism(
name='LVOC ControlMechanism',
features=[color_stim.input_state, word_stim.input_state],
# features={pnl.SHADOW_EXTERNAL_INPUTS: [color_stim, word_stim]},
objective_mechanism=pnl.ObjectiveMechanism(
name='LVOC ObjectiveMechanism',
monitor=[
task_decision.output_states[pnl.PROBABILITY_UPPER_THRESHOLD],
task_decision.output_states[pnl.PROBABILITY_LOWER_THRESHOLD],
reward
],
# monitored_output_states=[task_decision, reward],
function=objective_function),
agent_rep=pnl.RegressionCFA(
update_weights=pnl.BayesGLM(
mu_0=-0.17,
sigma_0=9.0909), # -0.17, 9.0909 precision = 0.11; 1/p = v
prediction_terms=[pnl.PV.FC, pnl.PV.COST]),
function=pnl.GradientOptimization(
convergence_criterion=pnl.VALUE,
convergence_threshold=0.001, #0.001
step_size=1, #1
annealing_function=lambda x, y: x / np.sqrt(y)
# direction=pnl.ASCENT
cost_options=[pnl.CostFunctions.INTENSITY, pnl.CostFunctions.ADJUSTMENT],
intensity_cost_function=pnl.Exponential(scale=1, rate=1),
# adjustment_cost_function=pnl.Exponential(scale=1, rate=1, offset=-1),
#offset = -1
allocation_samples=signalSearchRange)
Umemoto_comp.add_model_based_optimizer(
optimizer=pnl.OptimizationControlMechanism(
agent_rep=Umemoto_comp,
state_features=[
Target_Stim.input_port, Distractor_Stim.input_port,
Reward.input_port
],
state_feature_function=pnl.AdaptiveIntegrator(rate=1.0),
objective_mechanism=pnl.ObjectiveMechanism(monitor_for_control=[
Reward,
(Decision.output_ports[pnl.PROBABILITY_UPPER_THRESHOLD], 1, -1)
], ),
function=pnl.GridSearch(save_values=True),
control_signals=[
Target_Rep_Control_Signal, Distractor_Rep_Control_Signal
],
compute_reconfiguration_cost=pnl.Distance(
metric=pnl.EUCLIDEAN) # 0.8046),
))
Umemoto_comp.enable_model_based_optimizer = True
# Umemoto_comp.model_based_optimizer.set_log_conditions('value')
print(Target_Rep.loggable_items)
nTrials = 3
targetFeatures = [w_t]
flankerFeatures_inc = [w_d]
c_v_FitzHughNagumo=1.0,
d_v_FitzHughNagumo=0.0,
e_v_FitzHughNagumo=-1.0,
f_v_FitzHughNagumo=1.0,
a_w_FitzHughNagumo=1.0,
b_w_FitzHughNagumo=-1.0,
c_w_FitzHughNagumo=0.0,
t_0_FitzHughNagumo=0.0,
base_level_gain=G, # Additionally, we set the parameters k and G to compute the gain equation
scaling_factor_gain=k,
initial_v_FitzHughNagumo=initial_v, # Initialize v
initial_w_FitzHughNagumo=initial_w, # Initialize w
objective_mechanism=pnl.ObjectiveMechanism(
function=psyneulink.core.components.functions.transferfunctions.Linear,
monitored_output_states=[(
decision_layer, # Project output of T1 and T2 but not distractor from decision layer to LC
np.array([[lcwt], [lcwt], [0.0]])
)],
name='Combine values'
),
modulated_mechanisms=[decision_layer, response_layer], # Modulate gain of decision & response layers
name='LC'
)
# Log value of LC
LC.set_log_conditions('value')
# Set initial gain to G + k*initial_w, when the System runs the very first time,
# since the decison layer executes before the LC and hence needs one initial gain value to start with.
for output_state in LC.output_states:
output_state.value *= G + k * initial_w
# Construct Process
# Notes:
# The np.array specifies the matrix used as the Mapping Projection from input_layer to action_selection,
# which insures the left element of the input favors the left action (positive value of DDM decision variable),
# and the right element favors the right action (negative value of DDM decision variable)
# The learning argument specifies Reinforcement as the learning function for the Projection
comp = pnl.Composition()
p = comp.add_reinforcement_learning_pathway(
pathway=[input_layer, action_selection], learning_rate=0.5)
comp.add_controller(
pnl.OptimizationControlMechanism(
# pnl.ControlMechanism(
# objective_mechanism=True,
objective_mechanism=pnl.ObjectiveMechanism(monitor=action_selection),
control_signals=(pnl.LEARNING_RATE,
p.learning_components[pnl.LEARNING_MECHANISMS])))
# EXECUTION:
# Prints initial weight matrix for the Projection from the input_layer to the action_selection Mechanism
print('reward prediction weights: \n',
action_selection.input_port.path_afferents[0].matrix)
# Used by *call_before_trial* and *call_after_trial* to generate printouts.
# Note: should be replaced by use of logging functionality that has now been implemented.
def print_header(comp):
print("\n\n**** Time: ", comp.scheduler.get_clock(comp).simple_time)
time_constant_w_FHN=tau_u,
a_v_FHN=-1.0,
b_v_FHN=1.0,
c_v_FHN=1.0,
d_v_FHN=0.0,
e_v_FHN=-1.0,
f_v_FHN=1.0,
a_w_FHN=1.0,
b_w_FHN=-1.0,
c_w_FHN=0.0,
t_0_FHN=0.0,
initial_v_FHN=initial_v,
initial_w_FHN=initial_u,
objective_mechanism=pnl.ObjectiveMechanism(function=pnl.Linear,
monitored_output_states=[
(decision_layer, None, None,
np.array([[w_vX1], [0.0]]))
],
name='LC ObjectiveMechanism'),
modulated_mechanisms=[decision_layer, response_layer
], # Modulate gain of decision & response layers
name='LC')
task = pnl.System(processes=[decision_process])
# This displays a diagram of the System
task.show_graph(show_dimensions=True)
# Create Stimulus -----------------------------------------------------------------------------------------------------
# number of trials
trials = 1000
def test_evc_gratton(self):
# Stimulus Mechanisms
target_stim = pnl.TransferMechanism(name='Target Stimulus',
function=pnl.Linear(slope=0.3324))
flanker_stim = pnl.TransferMechanism(
name='Flanker Stimulus', function=pnl.Linear(slope=0.3545221843))
# Processing Mechanisms (Control)
Target_Rep = pnl.TransferMechanism(name='Target Representation')
Flanker_Rep = pnl.TransferMechanism(name='Flanker Representation')
# Processing Mechanism (Automatic)
Automatic_Component = pnl.TransferMechanism(name='Automatic Component')
# Decision Mechanism
Decision = pnl.DDM(name='Decision',
function=pnl.DriftDiffusionAnalytical(
drift_rate=(1.0),
threshold=(0.2645),
noise=(0.5),
starting_point=(0),
t0=0.15),
output_states=[
pnl.DECISION_VARIABLE, pnl.RESPONSE_TIME,
pnl.PROBABILITY_UPPER_THRESHOLD
])
# Outcome Mechanism
reward = pnl.TransferMechanism(name='reward')
# Pathways
target_control_pathway = [target_stim, Target_Rep, Decision]
flanker_control_pathway = [flanker_stim, Flanker_Rep, Decision]
target_automatic_pathway = [target_stim, Automatic_Component, Decision]
flanker_automatic_pathway = [
flanker_stim, Automatic_Component, Decision
]
pathways = [
target_control_pathway, flanker_control_pathway,
target_automatic_pathway, flanker_automatic_pathway
]
# Composition
evc_gratton = pnl.Composition(name="EVCGratton")
evc_gratton.add_node(Decision, required_roles=pnl.NodeRole.OUTPUT)
for path in pathways:
evc_gratton.add_linear_processing_pathway(path)
evc_gratton.add_node(reward, required_roles=pnl.NodeRole.OUTPUT)
# Control Signals
signalSearchRange = pnl.SampleSpec(start=1.0, stop=1.8, step=0.2)
target_rep_control_signal = pnl.ControlSignal(
projections=[(pnl.SLOPE, Target_Rep)],
function=pnl.Linear,
variable=1.0,
intensity_cost_function=pnl.Exponential(rate=0.8046),
allocation_samples=signalSearchRange)
flanker_rep_control_signal = pnl.ControlSignal(
projections=[(pnl.SLOPE, Flanker_Rep)],
function=pnl.Linear,
variable=1.0,
intensity_cost_function=pnl.Exponential(rate=0.8046),
allocation_samples=signalSearchRange)
objective_mech = pnl.ObjectiveMechanism(
function=pnl.LinearCombination(operation=pnl.PRODUCT),
monitor=[
reward,
(Decision.output_states[pnl.PROBABILITY_UPPER_THRESHOLD], 1,
-1)
])
# Model Based OCM (formerly controller)
evc_gratton.add_controller(controller=pnl.OptimizationControlMechanism(
agent_rep=evc_gratton,
features=[
target_stim.input_state, flanker_stim.input_state,
reward.input_state
],
feature_function=pnl.AdaptiveIntegrator(rate=1.0),
objective_mechanism=objective_mech,
function=pnl.GridSearch(),
control_signals=[
target_rep_control_signal, flanker_rep_control_signal
]))
evc_gratton.enable_controller = True
targetFeatures = [1, 1, 1]
flankerFeatures = [1, -1, 1]
rewardValues = [100, 100, 100]
stim_list_dict = {
target_stim: targetFeatures,
flanker_stim: flankerFeatures,
reward: rewardValues
}
evc_gratton.run(inputs=stim_list_dict)
expected_results_array = [[[0.32257752863413636], [0.9481940753514433],
[100.]],
[[0.42963678062444666],
[0.47661180945923376], [100.]],
[[0.300291026852769], [0.97089165101931],
[100.]]]
expected_sim_results_array = [
[[0.32257753], [0.94819408], [100.]],
[[0.31663196], [0.95508757], [100.]],
[[0.31093566], [0.96110142], [100.]],
[[0.30548947], [0.96633839], [100.]],
[[0.30029103], [0.97089165], [100.]],
[[0.3169957], [0.95468427], [100.]],
[[0.31128378], [0.9607499], [100.]],
[[0.30582202], [0.96603252], [100.]],
[[0.30060824], [0.9706259], [100.]],
[[0.29563774], [0.97461444], [100.]],
[[0.31163288], [0.96039533], [100.]],
[[0.30615555], [0.96572397], [100.]],
[[0.30092641], [0.97035779], [100.]],
[[0.2959409], [0.97438178], [100.]],
[[0.29119255], [0.97787196], [100.]],
[[0.30649004], [0.96541272], [100.]],
[[0.30124552], [0.97008732], [100.]],
[[0.29624499], [0.97414704], [100.]],
[[0.29148205], [0.97766847], [100.]],
[[0.28694892], [0.98071974], [100.]],
[[0.30156558], [0.96981445], [100.]],
[[0.29654999], [0.97391021], [100.]],
[[0.29177245], [0.97746315], [100.]],
[[0.28722523], [0.98054192], [100.]],
[[0.28289958], [0.98320731], [100.]],
[[0.42963678], [0.47661181], [100.]],
[[0.42846471], [0.43938586], [100.]],
[[0.42628176], [0.40282965], [100.]],
[[0.42314468], [0.36732207], [100.]],
[[0.41913221], [0.333198], [100.]],
[[0.42978939], [0.51176048], [100.]],
[[0.42959394], [0.47427693], [100.]],
[[0.4283576], [0.43708106], [100.]],
[[0.4261132], [0.40057958], [100.]],
[[0.422919], [0.36514906], [100.]],
[[0.42902209], [0.54679323], [100.]],
[[0.42980788], [0.50942101], [100.]],
[[0.42954704], [0.47194318], [100.]],
[[0.42824656], [0.43477897], [100.]],
[[0.42594094], [0.3983337], [100.]],
[[0.42735293], [0.58136855], [100.]],
[[0.42910149], [0.54447221], [100.]],
[[0.42982229], [0.50708112], [100.]],
[[0.42949608], [0.46961065], [100.]],
[[0.42813159], [0.43247968], [100.]],
[[0.42482049], [0.61516258], [100.]],
[[0.42749136], [0.57908829], [100.]],
[[0.42917687], [0.54214925], [100.]],
[[0.42983261], [0.50474093], [100.]],
[[0.42944107], [0.46727945], [100.]],
[[0.32257753], [0.94819408], [100.]],
[[0.31663196], [0.95508757], [100.]],
[[0.31093566], [0.96110142], [100.]],
[[0.30548947], [0.96633839], [100.]],
[[0.30029103], [0.97089165], [100.]],
[[0.3169957], [0.95468427], [100.]],
[[0.31128378], [0.9607499], [100.]],
[[0.30582202], [0.96603252], [100.]],
[[0.30060824], [0.9706259], [100.]],
[[0.29563774], [0.97461444], [100.]],
[[0.31163288], [0.96039533], [100.]],
[[0.30615555], [0.96572397], [100.]],
[[0.30092641], [0.97035779], [100.]],
[[0.2959409], [0.97438178], [100.]],
[[0.29119255], [0.97787196], [100.]],
[[0.30649004], [0.96541272], [100.]],
[[0.30124552], [0.97008732], [100.]],
[[0.29624499], [0.97414704], [100.]],
[[0.29148205], [0.97766847], [100.]],
[[0.28694892], [0.98071974], [100.]],
[[0.30156558], [0.96981445], [100.]],
[[0.29654999], [0.97391021], [100.]],
[[0.29177245], [0.97746315], [100.]],
[[0.28722523], [0.98054192], [100.]],
[[0.28289958], [0.98320731], [100.]],
]
for trial in range(len(evc_gratton.results)):
assert np.allclose(
expected_results_array[trial],
# Note: Skip decision variable OutputState
evc_gratton.results[trial][1:])
for simulation in range(len(evc_gratton.simulation_results)):
assert np.allclose(
expected_sim_results_array[simulation],
# Note: Skip decision variable OutputState
evc_gratton.simulation_results[simulation][1:])
def test_default_lc_control_mechanism(self):
G = 1.0
k = 0.5
starting_value_LC = 2.0
user_specified_gain = 1.0
A = pnl.TransferMechanism(
function=pnl.Logistic(gain=user_specified_gain), name='A')
B = pnl.TransferMechanism(
function=pnl.Logistic(gain=user_specified_gain), name='B')
# B.output_states[0].value *= 0.0 # Reset after init | Doesn't matter here b/c default var = zero, no intercept
LC = pnl.LCControlMechanism(modulated_mechanisms=[A, B],
base_level_gain=G,
scaling_factor_gain=k,
objective_mechanism=pnl.ObjectiveMechanism(
function=pnl.Linear,
monitor=[B],
name='LC ObjectiveMechanism'))
for output_state in LC.output_states:
output_state.value *= starting_value_LC
path = [A, B, LC]
S = pnl.Composition()
S.add_node(A, required_roles=pnl.NodeRole.INPUT)
S.add_linear_processing_pathway(pathway=path)
S.add_node(LC, required_roles=pnl.NodeRole.OUTPUT)
LC.reinitialize_when = pnl.Never()
gain_created_by_LC_output_state_1 = []
mod_gain_assigned_to_A = []
base_gain_assigned_to_A = []
mod_gain_assigned_to_B = []
base_gain_assigned_to_B = []
A_value = []
B_value = []
LC_value = []
def report_trial(system):
gain_created_by_LC_output_state_1.append(
LC.output_state.parameters.value.get(system)[0])
mod_gain_assigned_to_A.append(A.get_mod_gain(system))
mod_gain_assigned_to_B.append(B.get_mod_gain(system))
base_gain_assigned_to_A.append(A.function.parameters.gain.get())
base_gain_assigned_to_B.append(B.function.parameters.gain.get())
A_value.append(A.parameters.value.get(system))
B_value.append(B.parameters.value.get(system))
LC_value.append(LC.parameters.value.get(system))
result = S.run(inputs={A: [[1.0], [1.0], [1.0], [1.0], [1.0]]},
call_after_trial=functools.partial(report_trial, S))
# (1) First value of gain in mechanisms A and B must be whatever we hardcoded for LC starting value
assert mod_gain_assigned_to_A[0] == starting_value_LC
# (2) _gain should always be set to user-specified value
for i in range(5):
assert base_gain_assigned_to_A[i] == user_specified_gain
assert base_gain_assigned_to_B[i] == user_specified_gain
# (3) LC output on trial n becomes gain of A and B on trial n + 1
assert np.allclose(mod_gain_assigned_to_A[1:],
gain_created_by_LC_output_state_1[0:-1])
# (4) mechanisms A and B should always have the same gain values (b/c they are identical)
assert np.allclose(mod_gain_assigned_to_A, mod_gain_assigned_to_B)
def test_stability_flexibility_susan_and_sebastian(self):
# computeAccuracy(trialInformation)
# Inputs: trialInformation[0, 1, 2, 3]
# trialInformation[0] - Task Dimension : [0, 1] or [1, 0]
# trialInformation[1] - Stimulus Dimension: Congruent {[1, 1] or [-1, -1]} // Incongruent {[-1, 1] or [1, -1]}
# trialInformation[2] - Upper Threshold: Probability of DDM choosing upper bound
# trialInformation[3] - Lower Threshold: Probability of DDM choosing lower bound
def computeAccuracy(trialInformation):
# Unload contents of trialInformation
# Origin Node Inputs
taskInputs = trialInformation[0]
stimulusInputs = trialInformation[1]
# DDM Outputs
upperThreshold = trialInformation[2]
lowerThreshold = trialInformation[3]
# Keep Track of Accuracy
accuracy = []
# Beginning of Accuracy Calculation
colorTrial = (taskInputs[0] == 1)
motionTrial = (taskInputs[1] == 1)
# Based on the task dimension information, decide which response is "correct"
# Obtain accuracy probability from DDM thresholds in "correct" direction
if colorTrial:
if stimulusInputs[0] == 1:
accuracy.append(upperThreshold)
elif stimulusInputs[0] == -1:
accuracy.append(lowerThreshold)
if motionTrial:
if stimulusInputs[1] == 1:
accuracy.append(upperThreshold)
elif stimulusInputs[1] == -1:
accuracy.append(lowerThreshold)
# Accounts for initialization runs that have no variable input
if len(accuracy) == 0:
accuracy = [0]
# print("Accuracy: ", accuracy[0])
# print()
return [accuracy]
# BEGIN: Composition Construction
# Constants as defined in Musslick et al. 2018
tau = 0.9 # Time Constant
DRIFT = 1 # Drift Rate
STARTING_POINT = 0.0 # Starting Point
THRESHOLD = 0.0475 # Threshold
NOISE = 0.04 # Noise
T0 = 0.2 # T0
# Task Layer: [Color, Motion] {0, 1} Mutually Exclusive
# Origin Node
taskLayer = pnl.TransferMechanism(default_variable=[[0.0, 0.0]],
size=2,
function=pnl.Linear(slope=1,
intercept=0),
output_states=[pnl.RESULT],
name='Task Input [I1, I2]')
# Stimulus Layer: [Color Stimulus, Motion Stimulus]
# Origin Node
stimulusInfo = pnl.TransferMechanism(default_variable=[[0.0, 0.0]],
size=2,
function=pnl.Linear(slope=1,
intercept=0),
output_states=[pnl.RESULT],
name="Stimulus Input [S1, S2]")
# Activation Layer: [Color Activation, Motion Activation]
# Recurrent: Self Excitation, Mutual Inhibition
# Controlled: Gain Parameter
activation = pnl.RecurrentTransferMechanism(
default_variable=[[0.0, 0.0]],
function=pnl.Logistic(gain=1.0),
matrix=[[1.0, -1.0], [-1.0, 1.0]],
integrator_mode=True,
integrator_function=pnl.AdaptiveIntegrator(rate=(tau)),
initial_value=np.array([[0.0, 0.0]]),
output_states=[pnl.RESULT],
name='Task Activations [Act 1, Act 2]')
# Hadamard product of Activation and Stimulus Information
nonAutomaticComponent = pnl.TransferMechanism(
default_variable=[[0.0, 0.0]],
size=2,
function=pnl.Linear(slope=1, intercept=0),
input_states=pnl.InputState(combine=pnl.PRODUCT),
output_states=[pnl.RESULT],
name='Non-Automatic Component [S1*Activity1, S2*Activity2]')
# Summation of nonAutomatic and Automatic Components
ddmCombination = pnl.TransferMechanism(
size=1,
function=pnl.Linear(slope=1, intercept=0),
input_states=pnl.InputState(combine=pnl.SUM),
output_states=[pnl.RESULT],
name="Drift = (S1 + S2) + (S1*Activity1 + S2*Activity2)")
decisionMaker = pnl.DDM(function=pnl.DriftDiffusionAnalytical(
drift_rate=DRIFT,
starting_point=STARTING_POINT,
threshold=THRESHOLD,
noise=NOISE,
t0=T0),
output_states=[
pnl.DECISION_VARIABLE, pnl.RESPONSE_TIME,
pnl.PROBABILITY_UPPER_THRESHOLD,
pnl.PROBABILITY_LOWER_THRESHOLD
],
name='DDM')
taskLayer.set_log_conditions([pnl.RESULT])
stimulusInfo.set_log_conditions([pnl.RESULT])
activation.set_log_conditions([pnl.RESULT, "mod_gain"])
nonAutomaticComponent.set_log_conditions([pnl.RESULT])
ddmCombination.set_log_conditions([pnl.RESULT])
decisionMaker.set_log_conditions([
pnl.PROBABILITY_UPPER_THRESHOLD, pnl.PROBABILITY_LOWER_THRESHOLD,
pnl.DECISION_VARIABLE, pnl.RESPONSE_TIME
])
# Composition Creation
stabilityFlexibility = pnl.Composition(controller_mode=pnl.BEFORE)
# Node Creation
stabilityFlexibility.add_node(taskLayer)
stabilityFlexibility.add_node(activation)
stabilityFlexibility.add_node(nonAutomaticComponent)
stabilityFlexibility.add_node(stimulusInfo)
stabilityFlexibility.add_node(ddmCombination)
stabilityFlexibility.add_node(decisionMaker)
# Projection Creation
stabilityFlexibility.add_projection(sender=taskLayer,
receiver=activation)
stabilityFlexibility.add_projection(sender=activation,
receiver=nonAutomaticComponent)
stabilityFlexibility.add_projection(sender=stimulusInfo,
receiver=nonAutomaticComponent)
stabilityFlexibility.add_projection(sender=stimulusInfo,
receiver=ddmCombination)
stabilityFlexibility.add_projection(sender=nonAutomaticComponent,
receiver=ddmCombination)
stabilityFlexibility.add_projection(sender=ddmCombination,
receiver=decisionMaker)
# Beginning of Controller
# Grid Search Range
searchRange = pnl.SampleSpec(start=1.0, stop=1.9, num=10)
# Modulate the GAIN parameter from activation layer
# Initalize cost function as 0
signal = pnl.ControlSignal(
projections=[(pnl.GAIN, activation)],
function=pnl.Linear,
variable=1.0,
intensity_cost_function=pnl.Linear(slope=0.0),
allocation_samples=searchRange)
# Use the computeAccuracy function to obtain selection values
# Pass in 4 arguments whenever computeRewardRate is called
objectiveMechanism = pnl.ObjectiveMechanism(
monitor=[
taskLayer, stimulusInfo,
(pnl.PROBABILITY_UPPER_THRESHOLD, decisionMaker),
(pnl.PROBABILITY_LOWER_THRESHOLD, decisionMaker)
],
function=computeAccuracy,
name="Controller Objective Mechanism")
# Sets trial history for simulations over specified signal search parameters
metaController = pnl.OptimizationControlMechanism(
agent_rep=stabilityFlexibility,
features=[taskLayer.input_state, stimulusInfo.input_state],
feature_function=pnl.Buffer(history=10),
name="Controller",
objective_mechanism=objectiveMechanism,
function=pnl.GridSearch(),
control_signals=[signal])
stabilityFlexibility.add_controller(metaController)
stabilityFlexibility.enable_controller = True
# stabilityFlexibility.model_based_optimizer_mode = pnl.BEFORE
for i in range(1, len(stabilityFlexibility.controller.input_states)):
stabilityFlexibility.controller.input_states[
i].function.reinitialize()
# Origin Node Inputs
taskTrain = [[1, 0], [0, 1], [1, 0], [0, 1]]
stimulusTrain = [[1, -1], [-1, 1], [1, -1], [-1, 1]]
inputs = {taskLayer: taskTrain, stimulusInfo: stimulusTrain}
stabilityFlexibility.run(inputs)
decision_process = pnl.Process(default_variable=[0, 0],
pathway=[input_layer,
action_selection],
learning=pnl.LearningProjection(learning_function=psyneulink.core.components.functions
.learningfunctions.Reinforcement(
learning_rate=0.03)), # if learning rate set to .3 output state values annealing to [0., 0.]
# which leads to error in reward function
target=0
)
print('reward prediction weights: \n', action_selection.input_state.path_afferents[0].matrix)
print('target_mechanism weights: \n', action_selection.output_state.efferents[0].matrix)
conflict_process = pnl.Process(pathway=[action_selection, conflicts])
LC_NE = pnl.LCControlMechanism(objective_mechanism=pnl.ObjectiveMechanism(monitored_output_states=[action_selection],
name='LC-NE ObjectiveMech'),
modulated_mechanisms=[action_selection],
integration_method='EULER',
initial_w_FitzHughNagumo=initial_u,
initial_v_FitzHughNagumo=initial_v,
time_step_size_FitzHughNagumo=dt,
t_0_FitzHughNagumo=0.0,
a_v_FitzHughNagumo=-1.0,
b_v_FitzHughNagumo=1.0,
c_v_FitzHughNagumo=1.0,
d_v_FitzHughNagumo=0.0,
e_v_FitzHughNagumo=-1.0,
f_v_FitzHughNagumo=1.0,
time_constant_v_FitzHughNagumo=tau_v,
a_w_FitzHughNagumo=1.0,
b_w_FitzHughNagumo=-1.0,
comp = pnl.Composition(name='comp')
comp.add_node(reward, required_roles=[pnl.NodeRole.OUTPUT])
comp.add_node(Decision, required_roles=[pnl.NodeRole.OUTPUT])
task_execution_pathway = [Input, pnl.IDENTITY_MATRIX, Decision]
comp.add_linear_processing_pathway(task_execution_pathway)
comp.add_controller(
controller=pnl.OptimizationControlMechanism(
agent_rep=comp,
features=[Input.input_port, reward.input_port],
feature_function=pnl.AdaptiveIntegrator(rate=0.5),
objective_mechanism=pnl.ObjectiveMechanism(
function=pnl.LinearCombination(operation=pnl.PRODUCT),
monitor=[
reward,
Decision.output_ports[pnl.PROBABILITY_UPPER_THRESHOLD],
(Decision.output_ports[pnl.RESPONSE_TIME], -1, 1),
],
),
function=pnl.GridSearch,
control=[
{
pnl.PROJECTIONS: ('drift_rate', Decision),
pnl.ALLOCATION_SAMPLES: np.arange(0.1, 1.01, 0.3),
},
{
pnl.PROJECTIONS: (pnl.THRESHOLD, Decision),
pnl.ALLOCATION_SAMPLES: np.arange(0.1, 1.01, 0.3),
},
],
)
time_constant_w_FitzHughNagumo=tau_u,
a_v_FitzHughNagumo=-1.0,
b_v_FitzHughNagumo=1.0,
c_v_FitzHughNagumo=1.0,
d_v_FitzHughNagumo=0.0,
e_v_FitzHughNagumo=-1.0,
f_v_FitzHughNagumo=1.0,
a_w_FitzHughNagumo=1.0,
b_w_FitzHughNagumo=-1.0,
c_w_FitzHughNagumo=0.0,
t_0_FitzHughNagumo=0.0,
initial_v_FitzHughNagumo=initial_v,
initial_w_FitzHughNagumo=initial_u,
objective_mechanism=pnl.ObjectiveMechanism(
function=psyneulink.core.components.functions.transferfunctions.Linear,
monitored_output_ports=[(decision_layer, None, None,
np.array([[w_vX1], [0.0]]))],
name='LC ObjectiveMechanism'),
modulated_mechanisms=[decision_layer, response_layer
], # Modulate gain of decision & response layers
name='LC')
LC_process = pnl.Process(pathway=[LC])
task = pnl.System(processes=[decision_process, LC_process],
reinitialize_mechanisms_when=pnl.Never())
# This displays a diagram of the System
# task.show_graph()
# Create Stimulus -----------------------------------------------------------------------------------------------------
)
comp = pnl.Composition(name="evc")
comp.add_node(Reward, required_roles=[pnl.NodeRole.TERMINAL])
comp.add_node(Decision, required_roles=[pnl.NodeRole.TERMINAL])
task_execution_pathway = [Input, pnl.IDENTITY_MATRIX, Decision]
comp.add_linear_processing_pathway(task_execution_pathway)
ocm = pnl.OptimizationControlMechanism(
state_features=[Input, Reward],
state_feature_function=pnl.AdaptiveIntegrator(rate=0.5),
agent_rep=comp,
# function=pnl.GaussianProcessOptimization,
function=pnl.GridSearch,
control_signals=[("drift_rate", Decision), ("threshold", Decision)],
objective_mechanism=pnl.ObjectiveMechanism(monitor_for_control=[
Reward, Decision.PROBABILITY_UPPER_THRESHOLD,
(Decision.RESPONSE_TIME, -1, 1)
]))
comp.add_controller(controller=ocm)
comp.enable_controller = True
# Stimuli
comp._analyze_graph()
stim_list_dict = {Input: [0.5, 0.123], Reward: [20, 20]}
# print("- - - - - - - - RUN - - - - - - - -")
comp.show_graph()
# print (comp.run(inputs=stim_list_dict))
stabilityFlexibility.add_projection(sender=ddmCombination,
receiver=decisionMaker)
# beginning of Controller
search_range = pnl.SampleSpec(start=0.1, stop=0.3, num=3)
signal = pnl.ControlSignal(modulates=[(pnl.GAIN, activation)],
function=pnl.Linear,
variable=1.0,
intensity_cost_function=pnl.Linear(slope=0.),
allocation_samples=search_range)
objective_mech = pnl.ObjectiveMechanism(monitor=[
inputLayer, stimulusInfo, (pnl.PROBABILITY_UPPER_THRESHOLD, decisionMaker),
(pnl.PROBABILITY_LOWER_THRESHOLD, decisionMaker)
],
function=computeAccuracy)
meta_controller = pnl.OptimizationControlMechanism(
agent_rep=stabilityFlexibility,
features=[inputLayer.input_port, stimulusInfo.input_port],
feature_function=pnl.Buffer(history=3),
objective_mechanism=objective_mech,
function=pnl.GridSearch(),
control_signals=[signal])
inputs = {inputLayer: INPUT, stimulusInfo: stimulusInput}
stabilityFlexibility.add_controller(meta_controller)
stabilityFlexibility.enable_model_based_optimizer = True
def test_default_lc_control_mechanism(self, benchmark, mode):
G = 1.0
k = 0.5
starting_value_LC = 2.0
user_specified_gain = 1.0
A = pnl.TransferMechanism(function=pnl.Logistic(gain=user_specified_gain), name='A')
B = pnl.TransferMechanism(function=pnl.Logistic(gain=user_specified_gain), name='B')
# B.output_states[0].value *= 0.0 # Reset after init | Doesn't matter here b/c default var = zero, no intercept
LC = pnl.LCControlMechanism(
modulated_mechanisms=[A, B],
base_level_gain=G,
scaling_factor_gain=k,
objective_mechanism=pnl.ObjectiveMechanism(
function=pnl.Linear,
monitored_output_states=[B],
name='LC ObjectiveMechanism'
)
)
for output_state in LC.output_states:
output_state.value *= starting_value_LC
P = pnl.Process(pathway=[A, B, LC])
S = pnl.System(processes=[P])
LC.reinitialize_when = pnl.Never()
# THIS CURRENTLY DOES NOT WORK:
# P = pnl.Process(pathway=[A, B])
# P2 = pnl.Process(pathway=[LC])
# S = pnl.System(processes=[P, P2])
# S.show_graph()
gain_created_by_LC_output_state_1 = []
mod_gain_assigned_to_A = []
base_gain_assigned_to_A = []
mod_gain_assigned_to_B = []
base_gain_assigned_to_B = []
def report_trial():
gain_created_by_LC_output_state_1.append(LC.output_states[0].value[0])
mod_gain_assigned_to_A.append(A.mod_gain)
mod_gain_assigned_to_B.append(B.mod_gain)
base_gain_assigned_to_A.append(A.function_object.gain)
base_gain_assigned_to_B.append(B.function_object.gain)
benchmark(S.run, inputs={A: [[1.0], [1.0], [1.0], [1.0], [1.0]]},
call_after_trial=report_trial)
# (1) First value of gain in mechanisms A and B must be whatever we hardcoded for LC starting value
assert mod_gain_assigned_to_A[0] == starting_value_LC
# (2) _gain should always be set to user-specified value
for i in range(5):
assert base_gain_assigned_to_A[i] == user_specified_gain
assert base_gain_assigned_to_B[i] == user_specified_gain
# (3) LC output on trial n becomes gain of A and B on trial n + 1
assert np.allclose(mod_gain_assigned_to_A[1:], gain_created_by_LC_output_state_1[0:-1])
# (4) mechanisms A and B should always have the same gain values (b/c they are identical)
assert np.allclose(mod_gain_assigned_to_A, mod_gain_assigned_to_B)
target_rep_control_signal = pnl.ControlSignal(
projections=[(pnl.SLOPE, Target_Rep)],
variable=1.0,
intensity_cost_function=pnl.Exponential(rate=0.8046),
allocation_samples=signalSearchRange)
flanker_rep_control_signal = pnl.ControlSignal(
projections=[(pnl.SLOPE, Flanker_Rep)],
variable=1.0,
intensity_cost_function=pnl.Exponential(rate=0.8046),
allocation_samples=signalSearchRange)
objective_mech = pnl.ObjectiveMechanism(
function=pnl.LinearCombination(operation=pnl.PRODUCT),
monitor=[
reward, (Decision.output_ports[pnl.PROBABILITY_UPPER_THRESHOLD], 1, -1)
])
# Model Based OCM (formerly controller)
evc_gratton.add_controller(controller=pnl.OptimizationControlMechanism(
agent_rep=evc_gratton,
features=[
target_stim.input_port, flanker_stim.input_port, reward.input_port
],
feature_function=pnl.AdaptiveIntegrator(rate=1.0),
objective_mechanism=objective_mech,
function=pnl.GridSearch(),
control_signals=[target_rep_control_signal, flanker_rep_control_signal]))
evc_gratton.show_graph(show_controller=True)
task = pnl.LCAMechanism(name="TASK", size=2, initial_value=[0.5, 0.5])
task_color_wts = np.array([[4, 4], [0, 0]])
task_word_wts = np.array([[0, 0], [4, 4]])
task_color_pathway = [task_input, task, task_color_wts, color_hidden]
task_word_pathway = [task_input, task, task_word_wts, word_hidden]
# Construct the decision pathway:
decision = pnl.DDM(name="DECISION", input_format=pnl.ARRAY)
decision_pathway = [output, decision]
# Construct control mechanism
control = pnl.ControlMechanism(
name="CONTROL",
objective_mechanism=pnl.ObjectiveMechanism(
name="Conflict Monitor",
function=pnl.Energy(size=2, matrix=[[0, -2.5], [-2.5, 0]]),
monitor=output,
),
default_allocation=[0.5],
control_signals=[(pnl.GAIN, task)],
)
# Construct the Composition:
Stroop_model = pnl.Composition(name="Stroop_model")
Stroop_model.add_linear_processing_pathway(color_pathway)
Stroop_model.add_linear_processing_pathway(word_pathway)
Stroop_model.add_linear_processing_pathway(task_color_pathway)
Stroop_model.add_linear_processing_pathway(task_word_pathway)
Stroop_model.add_linear_processing_pathway(decision_pathway)
Stroop_model.add_controller(control)
# set up structure of inner comp
icomp.add_node(ia, required_roles=pnl.NodeRole.INPUT)
icomp.add_node(ib, required_roles=pnl.NodeRole.INPUT)
icomp.add_node(ic, required_roles=pnl.NodeRole.OUTPUT)
icomp.add_projection(pnl.MappingProjection(), sender=ia, receiver=ic)
icomp.add_projection(pnl.MappingProjection(), sender=ib, receiver=ic)
# set up inner comp controller and add to comp
icomp.add_controller(
pnl.OptimizationControlMechanism(
agent_rep=icomp,
features=[ia.input_port, ib.input_port],
name="Controller",
objective_mechanism=pnl.ObjectiveMechanism(
monitor=ic.output_port,
function=pnl.SimpleIntegrator,
name="iController Objective Mechanism"
),
function=pnl.GridSearch(direction=pnl.MAXIMIZE),
control_signals=[pnl.ControlSignal(projections=[(pnl.SLOPE, ia)],
variable=1.0,
intensity_cost_function=pnl.Linear(slope=0.0),
allocation_samples=pnl.SampleSpec(start=1.0,
stop=5.0,
num=5))])
)
# instantiate outer comp
ocomp = pnl.Composition(name='ocomp', controller_mode=pnl.BEFORE)
# setup structure for outer comp
c.add_node(word_task, required_roles=pnl.NodeRole.ORIGIN)
c.add_node(reward)
c.add_node(task_decision)
c.add_projection(sender=color_task, receiver=task_decision)
c.add_projection(sender=word_task, receiver=task_decision)
lvoc = pnl.OptimizationControlMechanism(
name='LVOC ControlMechanism',
features=[color_stim.input_port, word_stim.input_port],
# features={pnl.SHADOW_EXTERNAL_INPUTS: [color_stim, word_stim]},
# computes value of processing, reward received
objective_mechanism=pnl.ObjectiveMechanism(
name='LVOC ObjectiveMechanism',
monitor=[
task_decision.output_ports[pnl.PROBABILITY_UPPER_THRESHOLD],
task_decision.output_ports[pnl.PROBABILITY_LOWER_THRESHOLD],
reward, task_decision.output_ports[pnl.RESPONSE_TIME]
],
function=objective_function),
# posterior weight distribution
agent_rep=pnl.RegressionCFA(
# update_weights=pnl.BayesGLM(mu_0=-0.17, sigma_0=0.11), #sigma_0=math.sqrt(0.11))
update_weights=pnl.BayesGLM(
mu_0=-0.17, sigma_0=0.0000000000000001), #sigma_0=math.sqrt(0.11))
# update_weights=pnl.BayesGLM(mu_0=+0.17, sigma_0=0.11), #sigma_0=math.sqrt(0.11))
prediction_terms=[pnl.PV.C, pnl.PV.FC, pnl.PV.FF, pnl.PV.COST]),
# sample control allocs, and return best
# evaluate() computes outcome (obj mech) - costs given state (features) and sample ctrl alloc
function=pnl.GradientOptimization(
convergence_criterion=pnl.VALUE,
convergence_threshold=0.001,
