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,
# function=pnl.OptimizationFunction,
# control_signals={'COLOR CONTROL':[(pnl.SLOPE, color_task),
# ('color_control', word_task)]}
# control_signals={pnl.NAME:'COLOR CONTROL',
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
),
control_signals=pnl.ControlSignal(
projections=[(pnl.SLOPE, color_task), ('color_control', word_task)],
# function=pnl.ReLU,
function=pnl.Logistic,
cost_options=[
pnl.ControlSignalCosts.INTENSITY, pnl.ControlSignalCosts.ADJUSTMENT
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,
step_size=2, #1
# Note: Falk used 10 in the denom below, but indexed sample numbers from 1;
# but sample_num passed to _follow_gradient is indexed from 0, so use 11 below
annealing_function=lambda x, y: x / np.sqrt(11 + y),
max_iterations=100
# save_samples=True,
# save_values=True,
# direction=pnl.ASCENT