Two methods for optimising cognitive model parameters

Two methods for optimising
cognitive model parameters
David Peebles
—
9th August 2016

The problem
Searching parameter spaces can be laborious and time
consuming – particularly if done iteratively by hand
It’s difﬁcult to know if you have an optimal solution
Worse when model has several interacting parameters, is
non-differentiable, non-continuous, non-linear, stochastic, or
has many local optima.
Guilt and shame < boredom and frustration.

One solution, two methods
Use a population of models to explore the parameter space.
Differential evolution (DE)
Genetic algorithm for multidimensional real-valued functions.
Ideal for optimising models with relatively few parameters or
short run times on a single computer.
High-throughput computing with HTCondor
Open source software for managing job scheduling on
computer clusters or cycle scavenging idle computers.
Ideal for larger, more complex models or when simulating the
behaviour of many participants.
Used by many universities and easy to set up on local
networks (or so I’m told).

Differential evolution
An evolutionary strategy for real numbers used for
multidimensional numerical optimisation [2, 3].
Attractions:
- Simplicity of the algorithm
- Only three control parameters
- Fast, accurate, and robust performance
- Wide applicability
Control parameters:
NP the population size
F a scale factor applied to the mutation process
Cr a constant that regulates the crossover process

The DE algorithm in a nutshell
DE applies repeated cycles of mutation, recombination, and
selection on an initial, randomly generated population of vectors to
create a single vector that produces the best solution to a problem.
RecombinationInitialisation Mutation Selection

Initialisation
Create a population of real-valued vectors – one dimension
for each of the model parameters to be optimised.
Initialise vectors with uniformly distributed random numbers
within maximum and minimum bounds set for each dimension.
(:rt −5.0 −0.01) (:lf 0.1 1.5) (:ans 0.1 0.8)
To create the next population of vectors, each vector i in the
current population is selected in sequence, designated as the
target vector, and subjected to a competitive process. . .

Mutation
Mutation is the weighted difference between two vectors in the
current population.
A mutated donor vector is created by randomly selecting three
unique vectors, j, k and l from the population and adding the
difference between j and k (scaled by the F parameter) to l.
}=
j
k
l
}=
Donor vector
Sum
(weighted by F)
Difference
This ensures that differences in the scale and sensitivity of
different vector parameters are taken into account and that
the search space is explored equally on all dimensions.

Recombination
Cross the donor vector with the target vector to create the trial
vector to allow successful solutions from the previous
generation to be incorporated into the trial vector.
}=
Trial vector
Crossover
Donor
vector
vector
Target
Achieved by series of Bernoulli trials which determine for each
dimension which parent will donate its value.
Moderated by the crossover rate parameter Cr (0 ≤ Cr ≤ 1.).

Selection
Run the model with the parameter values from the trial vector.
If fitnesstrial ≥ fitnesstarget replace target vector with trial vector
in next generation, else retain target vector.
Repeat for each vector in the current population until the next
population is created and the evolutionary process continues
for a user-defined number of cycles.
Retain vector with highest fitness in each population.
Winning vector in the final population is considered the best
solution to the problem.

Paired Associate model example
1. Load (a) ACT-R, (b) DE code file, and (c) the fan effect model.
2. Modify the output function to just return one correlation.
3. Define a function to run the task suppressing the model output
and any warnings resulting from poor parameter choices.
4. Specify the three parameters to adjust with their ranges and
use the default DE configuration.
(defun output-data (data n)
(let ((probability (mapcar (lambda (x) (/ (first x) n)) data))
(latency (mapcar (lambda (x) (/ (or (second x) 0) n)) data)))
;;(correlation latency *paired-latencies*)
(correlation probability *paired-probability*)))
(defun run-paired-with-no-output ()
(let ((*standard-output* (make-string-output-stream)))
(suppress-warnings (paired-experiment 100))))
(optim ’run-paired-with-no-output ’((:rt -5.0 -0.01) (:lf 0.1 1.5) (:ans 0.1 0.8)))

HTCondor
DE useful for models with relatively few parameters or short
run times on a single computer.
DE too slow for large, complex models or when simulating the
behaviour of many participants.
HTCondor Pool
HPC Resources
Cloud Resources
HTCondor
Job
Manager
Users
Run a population of models over a network using HTCondor.
HTCondor – open source, cross-platform software system for
managing and scheduling tasks over computer networks [1].

Running ACT-R models on HTCondor
1. Modify your model to allocate random values to parameters.
2. Create a submit description file specifying job details
(executable, platform, the model files, commands etc.).
requirements =(OpSys == “WINNT61” && Arch == “INTEL”)
(OpSys == “WINDOWS” && Arch == “INTEL”)
(OpSys == “WINDOWS” && Arch == “X86 64”))
executable = actr-s-64.exe
arguments = “-l ’paired.lisp’ -e ’(collect-data 20)’ -e ’(quit)’ ”
transfer executable = ALWAYS
transfer input files = paired.lisp
output = out.stdout.$(Cluster).$(Process)
error = out.err.$(Cluster).$(Process)
log = out.clog.$(Cluster).$(Process)
queue 100

Random number generation
A potential problem arises with how the pseudorandom
number generator (PRNG) seed is set.
Each model instance must be initialised with a unique seed.
Many Lisps (and the ACT-R random module) use the
computer’s clock to generate seed. Problem if different
instances are given the same clock time.
To avoid this, use an ACT-R PRNG that doesn’t using the
system clock, uses a random number generated from CCL
(the Lisp used to create the MS Windows standalone ACT-R).
This method is employed in a new ACT-R
create-random-module which is loaded prior to loading the
model ﬁle.

Why not use MindModeling@Home?
NOT a toothbrush thing
MAYBE a lack of knowledge thing
MAYBE a World Community Grid thing (Zika, cancer cures).
Some models are small/trivial
Speed, convenience and ﬂexibility

Try them out
Code and instructions for using both methods using the paired
associate model from tutorial unit 4.
DE: https://github.com/peebz/actr-paired-de
HTCondor: https://github.com/peebz/actr-paired-htc
Thanks as always to Dan Bothell for his patience and advice.

References I
M. J. Litzkow, M. Livny, and M. W. Mutka. ‘Condor—A Hunter
of Idle Workstations’. In: Proceedings of the 8th International
Conference on Distributed Computing Systems. IEEE. June
1988, pp. 104–111.
R. Storn and K. Price. Differential evolution: A simple and
efﬁcient adaptive scheme for global optimization over
continuous spaces. Tech. rep. TR-95-012. Berkeley, CA: ICSI
Berkeley, 1995.
R. Storn and K. Price. ‘Differential evolution: A simple and
efﬁcient heuristic for global optimization over continuous
spaces’. In: Journal of global optimization 11.4 (1997),
pp. 341–359.

Two methods for optimising cognitive model parameters

More Related Content

What's hot (20)

Similar to Two methods for optimising cognitive model parameters (20)

More from University of Huddersfield (12)

Recently uploaded (20)

Two methods for optimising cognitive model parameters