interface=wm_interface))
defaults = Construct(name=buffer("defaults"),
process=Constants(strengths=default_strengths))
acs = Structure(name=subsystem("acs"))
with acs:
# We include a flow_in construct to handle converting chunk activations
# from working memory into features that the ACS can understand.
Construct(name=flow_in("wm"),
process=TopDown(source=buffer("wm"), chunks=nacs_cdb))
Construct(name=features("main"),
process=MaxNodes(sources=[
buffer("acs_ctrl"),
flow_in("wm"),
buffer("defaults")
]))
# This terminus controls the working memory.
Construct(name=terminus("wm"),
process=ActionSelector(source=features("main"),
temperature=.01,
interface=alice.assets.wm_interface))
# This terminus controls the agent's speech actions.
alice = Structure(name=agent("alice"))
with alice:
stimulus = Construct(name=buffer("stimulus"), process=Stimulus())
nacs = Structure(name=subsystem("nacs"), assets=Assets(chunk_db=chunk_db))
with nacs:
# Although the entry point for the NACS are chunks, in this example we
# start with features as there are no constructs that initially
# activate chunks in the NACS activation cycle.
Construct(name=features("main"),
process=MaxNodes(sources=[buffer("stimulus")]))
Construct(name=flow_bt("main"),
process=BottomUp(source=features("main"),
chunks=nacs.assets.chunk_db))
Construct(name=chunks("main"),
process=MaxNodes(sources=[flow_bt("main")]))
# Termini
# In addition to introducting chunk extraction, this example
# demonstrates the use of two temrmini in one single subsytem. We
# include one terminus for the output of the top level and one for the
# bottom level.
stimulus = Construct(
name=buffer("stimulus"),
process=Stimulus()
)
nacs = Structure(
name=subsystem("nacs"),
)
with nacs:
Construct(
name=flow_in("lag"),
process=Lag(
source=features("main"),
max_lag=1
)
)
Construct(
name=features("main"),
process= MaxNodes(
sources=[
buffer("stimulus"),
flow_in("lag")
]
)
)
defaults = Construct(name=buffer("defaults"),
process=Constants(strengths=default_strengths))
# This simulation adds an entirely new subsystem to the model: the
# action-centered subystem, which handles action selection. We keep this
# ACS to a bare minimum.
acs = Structure(name=subsystem("acs"))
with acs:
# Assembly of the ACS is similar to the NACS, but features are the
# (primary) entry points for activations.
Construct(
name=features("main"),
process=MaxNodes(sources=[buffer("acs_ctrl"),
buffer("defaults")]))
# We define an action terminus in ACS for controlling flow gating in
# NACS. To do this, we make use of the ActionSelector emitter, which
# selects, for each command dimension in its client interface, a single
# value through boltzmann sampling, and forwards activations of any
# parameter features defined in the interface.
Construct(name=terminus("nacs"),
process=ActionSelector(source=features("main"),
interface=alice.assets.gate_interface,
temperature=0.01))
# Next, we set up the NACS, adding the `Gated` wrapper where necessary.
# of a stimulus buffer. In more sophisticated models, reinforcements may be
# generated by the Meta-Cognitive Subsystem.
reinforcement = Construct(name=buffer("reinforcement"), process=Stimulus())
acs = Structure(name=subsystem("acs"))
with acs:
# We use a simple repeater to relay the actions selected on the
# previous step back to the qnet.
Construct(name=flow_in("ext_actions_lag1"),
process=Repeater(source=terminus("ext_actions")))
Construct(name=features("in"),
process=MaxNodes(sources=[buffer("sensory")]))
# We construct the Q-Net just like any other Process instance and
# integrate it into the bottom level. Note that it is designated as a
# construct of type flow_bb. The particular Process class used here,
# SimpleQNet, will construct an MLP with two hidden layers containing 5
# nodes each. Weight updates occur at each step (i.e., training is
# online) through gradient descent with backpropagation.
# On each trial, the q-net outputs its Q values to drive action
# selection at the designated terminus. The Q values are squashed prior
# being output to ensure that they lie in [0, 1].
qnet = Construct(name=flow_bb("q_net"),
process=SimpleQNet(