The Future of ACT-R
Revisited
Fourth
Annual ACT-R WorkshopCarnegie Mellon University
August 5, 1997
Christian Lebiere
cl+@cmu.edu
Last Years Changes
Eliminate lists as built-in type
unbounded span?
chunks as cons cells
Eliminate generalized negation
exhaustive search? latency?
conflict resolution can handle it
Eliminate instantiation-specific PG-C
q, a, r, b only as (learned) numbers
can still introduce an arbitrary evaluation function through the :value parameter
Disallow bindings to nil?
Plus
New dependency-based analogy
Better syntax
More flexible, general mechanism
Conflict Resolution
Currently:
Instantiations generated in parallel and evaluated in order of increasing latency.
Problems:
Cutoff latency is NOT matching latency.
Assumes excessive parallelism.
Proposal:
Productions that match the goal are
instantiated sequentially.
Highest PG-C considered first.
No multiple instantiations.
Production latencies are cumulative.
Latency of retrieval failure is latency
of activation threshold (0.0 by default).
Q & A & R & B
Currently:
Q&A and R&B are the probability of success and cost of the action of the present production and all productions until ultimate success or failure, resp.
The value of a subgoal is the PG-C of the instantiation, with multiple subgoals getting half the value of the previous one.
Problems:
Need to estimate the success of the
matching part of the production.
Arbitrary termination points for success
and failure leads to improper credit
assignment and lack of generalization.
Subgoals are penalized multiple times.
Subgoal value gets exponentially low.
Q & A & R & B cont.
Proposal:
Q&A apply to the matching, action and
subgoals of a production.
R&B apply to future productions until
the current goal is popped.
When learning R&B, a goal fails if the
popping production is marked as :failure
or if the goal is popped by default.
Otherwise the goal succeeds.
When learning Q&A, a production fails if
the matching or action part fails or any
of its subgoals fail.
The value of a subgoal is equal to rG-b
of the creating instantiation. For nth of
many subgoals, divide rG-b by n.
For analogized productions,
estimate Q&A from the matching part.
Popping upon Failure
Currently:
The best instantiation considered is fired. If none is found, the system stops.
Problems:
Sometimes gets lost considering
instantiations with very negative PG-C.
Sometimes gets stuck in subgoals.
Proposal:
Pop the goal when no instantiation can
be found with a positive PG-C.
Return a :failure value to signal the
failure to parent goals.
Provides an estimate of the time to
create (memorize) a wme by subgoaling.
Replaces usual Give Up production.
Chunks
Currently:
Multiple copies of same chunk (goal) are created.
Problems:
Subgoals have to be repeatedly
recomputed because they cannot be
retrieved or rehearsed.
Proposal:
When a goal is popped, it is merged with
any identical wme, which is reinforced.
Does not apply to rhs creation or
modification of wmes.
Activation Noise
Currently:
New Gaussian noise of mean 0 and variance AN is added to each wme at each iteration.
Problems:
Too high: wide fluctuation from cycle to
cycle in ability to recall chunks.
Too low: not enough randomness in
which chunks are recalled or not.
Proposal:
Introduce a Permanent Activation Noise
added to a wmes base level at creation.
Defaults to NIL (current situation).
High PAN and low (T)AN provides for
randomness but consistency in recall.
P/TAN ~ Representation/Processing.
The ACT-R System
(p future
ACT-R 3.0
development environment
visual interface
analogy authoring tools
experimental paradigm software
& models models models
==>
integrate system components
integrate independent models
)