The workshop will be held over two days. The format will consist of:
o twenty minute talks with ten minute discussions,
o a number of invited talks,
o poster and discussion sessions, and
o dinners in town.
Talks will be clustered by approach so that researchers unfamiliar
with the various aspects of action-selection modeling will have an
opportunity to learn a new area. We intend to allow speakers to know
the speaking order well in advance so that they can coordinate their
talks to maximize content and minimize repetition.
Talks will be chosen from submitted papers. All papers will be peer
reviewed. The number of full papers accepted, as determined by review,
may exceed the number of talk slots available, in which case the
remainder of accepted papers will be offered a special full-paper
poster session. The maximum number of participants is limited
to 40. If there is room for participants without full papers, a second
call will be sent out in May for extended abstracts and ordinary
poster submissions.
Invited Speakers:
- Prof. Randall C. O'Reilly, University of Colorado Boulder
- Prof. Michael Laver, New York University
- Dr. Marius Usher (Reader), Birkbeck University of London
Schedule:
Schedule:
Paper #1
Cognition, Action Selection, and Inner
Rehearsal
Dr. Murray Shanahan
Reader in Computational Intelligence
Dept. Electrical and Electronic Engineering
Imperial College London
This paper presents a large-scale model of the
architecture of the mammalian brain, the core circuit of which carries
out inner rehearsal of interaction with the environment to realise
a form of cognitively mediated action selection. As it alternates
between broadcast to and competition between its component neural
assemblies, the core circuit exhibits an episodic dynamics suggestive
of cortical processing in discrete frames. The implemented architecture
is used to control a simulated robot, and a classic experimental
paradigm in which rats performed apparently goal-directed action
selection is emulated.
No keywords provided.
Back to top
Paper #2
How a biological decision network can
implement a statistically optimal test
Rafal Bogacz
Department of Computer Science
University of Bristol
Neurophysiological evidence due to Schall, Newsome and others indicates
that decision processes in certain cortical areas (e.g. FEF and
LIP) involve the integration of noisy evidence. Within this paradigm,
we ask which neuronal architectures and parameter values would allow an
animal to make the fastest and most accurate decisions. Since evolutionary
pressure promotes such optimality (e.g. in prey capture and predator
avoidance), it is plausible that biological decision networks realise
or approximate optimal performance. We consider a simple decision model
consisting of two populations of neurons integrating evidence in
support of two alternatives. We show that in order to implement the optimal
decision algorithm (sequential probability ratio test) the linearised
network must satisfy the following two constraints: (i) it must
accumulate the difference between evidence in support of each
alternative, as would be implemented by mutual inhibition between
the populations; and (ii) the strength of mutual inhibition must be
equal to the leak of activity from each population.
decision making, action selection, optimality, sequential probability
ratio test
Back to top
Paper #3
When and when not to use your subthalamic
nucleus: Lessons from a computational model of the basal ganglia
Michael J Frank
University of Colorado at Boulder
The basal ganglia (BG) are thought to coordinate response selection
processes by facilitating adaptive frontal motor commands while
suppressing those that are less adaptive. In previous work, a neural
network model of the BG accounted for cognitive deficits in
reinforcement learning and response selection associated with BG<
dopamine depletion in Parkinson's disease. Novel predictions
from this model have been subsequently confirmed in medicated and
non-medicated patients and in healthy participants taking low doses
of dopamine medications. Nevertheless, one clear limitation of the
model is in its omission of the subthalamic nucleus (STN), which
is a key BG structure that participates in both motor and cognitive
processes. Here I include the STN and show that by modulating {\em
when} a response is executed, it can reduce premature responding
and therefore have substantial effects on {\em which} response is
ultimately selected, particularly when there are multiple competing
responses. The model accurately captures the dynamics of activity
in various BG areas during response selection. Simulated
Parkinsonism, by depleting dopamine from the system, results in
emergent oscillatory activity in STN and BG output structures, which
has been linked with Parkinson's tremor. Finally, the model accounts
for the beneficial effects of STN lesions on these oscillations, but
suggests that this benefit may come at the expense of impaired
decision making.
Neural network models; basal ganglia; subthalamic
nucleus; dopamine; response selection
Back to top
Paper #4
Predicting violations of transitivity
when choices involve fixed or variable delays to food.
Prof. Alasdair I. Houston
Centre for Behavioural Biology
School of Biological Sciences
University of Bristol
Incentive theory, an established model of behaviour, predicts how
animals should choose between alternatives that differ in the amount
of food delivered, and the delay until it is delivered. Choice
behaviour in such situations may be "irrational", in that it fails
to satisfy Strong Stochastic Transitivity (SST). We apply incentive theory
within the framework of the standard choice procedure from operant psychology
and show that choice behaviour does not satisfy substitutability,
and therefore SST does not hold. This occurs due to a change in
the context of choice, implicit in the change of experimental conditions
necessary to test SST. Our findings show similar patterns
to the results of experimental studies of choice behaviour in pigeons.
This agreement suggests that behavioural theories may provide insight
into other apparent departures from rational behaviour.
Rational choice, intransitivity, choice context
Back to top
Paper #5
Estimation of eye-pupil size during blink
by Support Vector Regression
Minoru NAKAYAMA
CRADLE(The Center for R&D of Educational Technology)
Tokyo Institute of Technology, Japan
Pupillography can be an index of mental activity
and
sleepiness, however blinks prevent its measurability as an
artifact. A method of estimation of pupil size from
pupillary changes during blinks was developed using a
support vector regression technique. Pupil responses for
changes in periodic brightness were prepared, and
appropriate pupil sizes for blinks were given as a set of
training data. The performance of the trained estimation
models were compared and an optimized model was obtained.
An examination of this revealed that its estimation
performance was better than that of the estimation method
using MLP. This development helps in the understanding of
the behaviour of pupillary change and blink action.
Eye-pupil, blink, estimation, Support Vector Regression
Back to top
Paper #6
Learning affordance concepts:
some seminal ideas
Manuela Viezzer
Thales Research and Technology
Nederland
Inspired by the pioneering work of J. J. Gibson, we
provide a workable characterisation of the notion of affordance
and we explore a possible architecture for an agent that is able to
autonomously acquire affordance concepts.
Affordances, cognitive architecture, concept formation
Back to top
Paper #7
Routine
Action: Combining Familiarity and Goal Orientedness
Nicolas Ruh
School of Psychology,
Birkbeck, University of London, Malet Street, London, WC1E 7HX, UK
Two current approaches
to modelling naturalistic sequential routine action selection differ
along two dimensions: (a) the number of systems required and (b)
the nature of the underlying task representation. We present findings
from a study that supports a combination of the two computational
accounts, namely a familiarity-dependent basic system, interfaced
with a higher-level supervisory system to bias it at crucial points
in a sequence. In order to elaborate this position, we explore a
connectionist reinforcement model of routine action that (a) learns
goal-directed action sequences through a combination of exploration
and exploitation, and (b) offers the prospect of being interfaced
with a supervisory system.
routine sequential action; dual systems; reinforcement
learning; recurrent neural networks
Back to top
Paper #8
Building agents to understand infant attachment
behaviour
Dean Petters
University of Birmingham,
School of Computer Science,
Birmingham B15 2TT, United Kingdom
This paper reports on autonomous agent simulations
of infant attachment behaviour. The behaviours simulated have been observed in home
environments and in a controlled laboratory procedure called the
Strange Situation Experiment. The Avoidant, Secure and Ambivalent styles
of behaviour seen in these studies are outlined, and then abstracted
to their core elements to act as a specification of requirements for the
simulation. A reactive agent architecture demonstrates that these
patterns of behaviour can be learnt from reinforcement signals without recourse
to deliberative mechanisms. Alternative designs based upon hybrid
architectures are also considered.
Attachment infant architecture reactive Bowlby
Back to top
Paper #9
Hierarchical
reactive planning: Where is its limit? ("Brom_MNAS_Gardener050331.pdf")
Cyril Brom
Charles University
in Prague
Faculty of Mathematics and Physics
Malostranské nám. 2/25, Prague, Czech Republic
Nowadays, hierarchical
reactive methods are very popular in the field of controlling complex
artificial intelligent agents. In this paper, we argue that they
can not cope with human-like behaviour.
We propose a detailed analysis of behaviour of a relatively simple
human-like artificial agent, an artificial gardener, whose action
selection model is based on hierarchical reactive planning with
rigid switching. It is shown that although the agent has no troubles
with survival in a complex and dynamic environment, it does not
behave believable in some situations. However, instead of rejection
the judged methodology, we propose how to extend it using certain
features of other action selection models.
Action selection,
human-like intelligent agent, hierarchical reactive planning
Back to top
Paper #10
Action
Selection in Subcortical Loops through Basal Ganglia
James C. Houk
Professor of Physiology at Northwestern University
Medical School in Chicago, IL, USA
Subcortical
loops through the basal ganglia and cerebellum form computationally
powerful distributed processing modules (DPMs). This paper relates
the computational features of a DPM's loop through the basal ganglia
to experimental results for two kinds of natural action selection.
First, functional imaging during a serial order recall task was
used to study human brain activity during the selection of sequential
actions from working memory. Second, microelectrode recordings from
monkeys trained in a step-tracking task were used to study the natural
selection of corrective submovements. Our DPM-based model assisted
in the interpretation of puzzling data from both of these experiments.
We come to posit that the many loops through the basal ganglia each
regulate the embodiment of pattern formation in a given area of
cerebral cortex. This operation serves to instantiate different
kinds of action (or thought) mediated by different areas of cerebral
cortex.
Basal ganglia, serial order, error correction,
functional imaging, neurophysiology
Back to top
Paper #11
Forced moves or good
tricks in design space? Great moments in the evolution of the neural substrate for action selection
Tony J. Prescott
Adaptive Behaviour Research
Group, University of Sheffield, UK.
This mini-review considers some
important landmarks in the evolution of animals, asking to what extent
specialised action selection mechanisms play a role in the
functional architecture of different nervous system plans, and looking
for 'forced moves' or 'good tricks' (Dennett, 1995) that could possibly
transfer to the design of control systems for mobile robots. A
key conclusion is that while cnidarians (e.g. jellyfish) appear to have
discovered some good tricks for the design of behaviour-based control
systems-lacking specialised selection mechanisms; the evolution
of bilaterality in platyhelminths (flatworms) may have forced
the evolution of a central ganglion (or 'archaic brain') whose main
function is to resolve conflicts between competing peripheral systems.
Whilst vertebrate nervous systems contain many interesting substrates
for selection it is likely that here too, the evolution of centralised
selection structures such as the basal ganglia and medial
reticular formation may have been a forced move due to the need
to limit connection costs as brains increased in size.
comparative neurobiology, jellyfish,
flatworms, vertebrates,
integration, architecture
Back to top
Paper #12
Reinforcement Learning of Stable Trajectory
for Quasi-Passive-Dynamic Walking
Kentarou HITOMI (1), Tomohiro SHIBATA (1,2), Yutaka
NAKAMURA (1), and Shin
ISHII (1)
(1)Nara Institute of Science and Techcnology
Takayama 8916-5, Ikoma, Nara 630-0192
(2)ATR Computational Neuroscience Laboratories
A class of biped locomotion called Passive Dynamic Walking (PDW)
has been recognized to be efficient in energy consumption and a key to understand
human walking. Although DW is sensitive to the initial condition
and disturbances, some studies of Quasi-PDW, which incorporates supplemental
actuators, have been reported to overcome the sensitivity. In this
article, we propose a reinforcement learning designed in particular for Quasi-PDW
walking. The keys of our approach are a reward function and a learning
scheme of a simple feedback controller, both of which utilize the
robot’s passive dynamics as much as possible. Computer simulations show
that the parameter in a Quasi-PDW controller is quickly tuned by our method,
and that the obtained controller is robust against variations in the slope
gradient.
2D biped, passive dynamic walk, reinforcement
learning, adaptive control
Back to top
Paper #13
Combining Action Selection Models with
a Five Factor Theory
Mark Witkowski
Intelligent Systems and Networks Group
Department of Electrical and Electronic Engineering
Imperial College
Exhibition Road, London, SW7 2BT, United Kingdom
This paper describes a unifying framework for five highly influential
but disparate theories (the five factors) of natural learning and behavioral
action selection. These theories are normally considered independently,
with their own experimental procedures and results. The framework builds
on a structure of connection types, propagation rules and learning rules,
which are used in combination to integrate results from each theory into
a whole. Exemplar experimental procedures will be used to discuss the areas
of genuine difference, and to identify areas where there is overlap
and where apparently disparate findings have a common source. The paper focuses
on predictive or anticipatory properties inherent in these action selection
and learning theories, and uses the Dynamic Expectancy Model and its
computer
implementation SRS/E as a mechanism to conduct this discussion.
Action selection, animal learning theory, prediction
and anticipatory learning, dynamic and static policy maps
Back to top
Paper #14
Goal and motor action selection using
a hippocampal and prefrontal
model
Nicolas Cuperlier, Philippe Gaussier, Philippe
Laroque, Mathias Quoy
ETIS-UMR 8051
Université de Cergy-Pontoise - ENSEA
6, Avenue du Ponceau \\95014 Cergy-Pontoise France\\
We
have developped a mobile robot control system based on hippocampus and prefrontal models. We
propose a model that rely on cognitive maps linking transition cells to
adress action selection problems encoutered in navigation tasks. The transition
cell links two locations with the integrated direction used.
We describe a simple neural mechanism able to learn and predict
these transitions to perform motivated path planning to reach goals. Furthermore,
we propose a dynamical system using a Neural Field for selecting
the final movement to perform. Simulations and robotics experiments
are carried out.
Action selection, hippocampus,
transition cells, dynamical system
Back to top
Paper #15
A computational model of reach decisions
in the primate cerebral cortex
Paul Cisek, Ph.D.
Department of physiology
University of Montreal
C.P. 6128 Succursale Centre-ville
Montreal QC H3C 3J7 Canada
phone: 514-343-6111 x4355
FAX: 514-343-2111
Neurophysiological evidence suggests that visually-guided
reaching movements are produced through "specification"
and "selection" processes that overlap both temporally
and anatomically (Cisek and Kalaska, 2005). Here, I present a formal
computational model which demonstrates how partial specification
of several potential movement directions, and the selection of the
correct movement, can occur in populations of directionally tuned
cells in a distributed cortical network including posterior parietal,
premotor, prefrontal, and primary motor cortex. The model reproduces
a large set of neurophysiological and psychophysical phenomena,
including the behavior of cortical cells during a reach decision
task and the spatial and temporal statistics of human reaching choices.
Computational Neuroscience, Decision-Making, Action
Selection, Neurophysiology, Cerebral Cortex
Back to top
Paper #16
Innate Planning Mechanisms
Sule Yildirim
Complex Adaptive Organically-Inspired Systems
Group
Department of Computer Science
The Norwegian University of Science and Technology, Trondheim, Norway
In
this paper, we will discuss whether there could be any means to
bridge the gap between the Symbolic
and Subsymbolic AI. One way to do this is to ask ourselves if human brain
executes any planning algorithms. We see that we have taken a series of
steps when we are done with planning in a situation. Taking a series
of steps during planning might be a result of the execution of an in-nate
planning algorithm. If we really are executing a planning al-gorithm, then
we believe, its function is very general and is to set the conditions which
will trigger a next step to take. A step to take might be the execution
of an IF rule as an example. IF rule execu-tions are not the only
steps to take while planning, however, for simplicity, they are assumed
as the only ones here. There is still not any neuro-scientific evidence
against the possibility that human mind incorporates an innate planning
algorithm that triggers the next rule to execute (the step to take)
yet. For that reason, in this paper we will investigate that possibility.
Action sequencing, action
associations, concept-action associations, emergence of
association mechanisms
Back to top
Paper #17
Building Plans for Household Tasks from
Distributed Knowledge
Rakesh Gupta
Honda Research Institute USA Inc
To accomplish a household task, an autonomous system needs
a plan with steps. It is desirable to derive this
plan dynamically instead of pre-coding it in the system. In this paper, we find
a plan by using common sense knowledge collected from volunteers over the
web through distributed knowledge capture techniques. This knowledge
consists of steps for executing common household tasks. We first pre-process the data with part-of-speech
(POS) tagging to identify the actions and objects in the steps in all available
plans for the task. We then determine the order of the steps to accomplish
the task using discriminative as well as generative models. For the
discriminative approach, we cluster the plans using hierarchical
agglomerative clustering and choose a plan from the biggest cluster. In the
contrasting approach, we make use of generative Markov chain techniques.
Using human judgments, we show that the generative model with
the first order Markov chain has the best performance. We also show that
environmental constraints can be incorporated in the generated plans.
Data Mining, Routine Actions, Task Model
Back to top
Paper #18
Prediction of the
Behavioural Strategy in a Goal-Oriented Task
Manuel A. Sanchez-Montanes
Escuela Politecnica Superior
Universidad Autonoma de Madrid (Spain)
Tim
C. Pearce
Department of Engineering
University of Leicester (UK)
A
common approach to understanding the behaviour of animals within their natural
habitat involves breaking down complex behavioral
sequences into simpler component behavioral
units. Here, we propose a theoretical
framework based upon Markov state descriptions
of animal behaviours which can be optimised in a virtual
environment. In this case, the sensory information
available to the animal is naturally characterized
by nonstationary statistics which depend on
the history of action selection. This framework
allows us to address many questions, such as how and
why behavioral units are arranged in a particular order,
the degree of optimality in the overall
behavioral strategy, and how much memory is required
to efficiently solve a behavioral task. To illustrate
our approach, we take a well characterised animal
behaviour -- chemotaxis in the moth -- which is performed
robustly in turbulent chemical plumes.
In this example the optimisation of a two-state
model in a virtual plume generates a behavioural
strategy, the units of which are directly comparable
to those seen in the animal -- 'cast' and 'surge'.
Finally, we discuss how our framework may be extended
to understand sensory integration and neuronal coding
within sensorimotor pathways that supports effective
natural action selection.
Chemical search, behaviour
prediction, behavioural quantification, partially observable environments,
neuroethology
Back to top
Paper #19
He/She-You-I Formalism:
A Heuristic Model of (En)Action to Make Decisions
Daniel Mellet-d'Huart
AFFILIATION AFPA DEAT / Université du Maine LIUM
This paper presents a heuristic model of (en)action that was developed
in order to
support the design of virtual environments for learning. The underlying
action model
is rooted in recent researches in neurosciences and based on a specific
conceptual
framework. It is organized as a three-pole system within the framework
of a "He/She-
You-I formalism". A "He/She Pole" deals with understanding,
simulation and
anticipation of action; a "You Pole" with making decision
and taking engagement
within action; and a "I Pole" with realizing action within
a particular environment.
Acting, action, enaction, model, design method, virtual reality
Back to top
Paper #20
Selection Actions and Making Decisions:
Lessons from AI Planning
Hector Geffner
ICREA & Universitat Pompeu Fabra, SPAIN
Humans and animals encounter a huge variety of problems
which they must solve using general methods. Even simple
problems, however, become computationally hard if the structure
of the problems is not recognized and exploited. Work in Artificial
Intelligence Planning and Problem Solving, faces a similar
difficulty,
and in recent years has led to well-founded and empirically tested
techniques for recognizing and exploiting structure, focusing the
search
for solutions, and in certain cases, bypassing the need to search
altogether. These techniques include the automatic derivation of
heuristic
functions, the use of limited but effective forms of inference,
and
the compilation of domains, all of which enable a general problem
solver
to `adapt' to the task at hand. In this paper, I present
the ideas
underlying these techniques, and argue for their relevance
to models
of natural intelligent behavior. The need for focusing the search
for
solutions has actually been recognized in a number of recent
works
concerned with natural intelligent behavior, where it has been related
to the role of emotions. In the paper, we address this issue as
well.
Search,
Heuristics, Emotion
Back to top
Paper #21
Simulation, Emotion and Information Processing: Computational
Investigations of the Regulative Role of Emotion in Adaptive
Behavior.
Joost Broekens (corresponding author) and Fons Verbeek
University of Leiden
Leiden Institute of Advanced Computer Science
Niels Bohrweg 1, 2333CA, Leiden, The Netherlands
Emotion plays an important role in thinking. Integration of emotion
at
multiple levels of information processing suggests that it contributes
to survival. We focus on the low-level regulatory influence of emotion
on information processing mechanisms and adaptive behavior, and
use a
computational reinforcement-learning model to study this. Our model
is
based upon Interactivism and the simulation hypothesis. We explain
how
the simulation hypothesis is embedded in the model. We measure emotion
as a function of the performance of the agent under interaction
with
its environment. The emotion is used as feedback to the simulation
strategy of the agent. Our main hypothesis tested is: a simulation
strategy that is dynamically adapted by emotional feedback provides
additional survival value to an agent compared to a static simulation
strategy. Experimental results illustrate that this hypothesis holds
true. Dynamic adaptation results in a learning performance that
at
least equals static simulation strategies. Additionally, it results
in
a major decrease of mental effort required for this performance.
This
is an important observation in the understanding of the evolutionary
plausibility of the simulation hypothesis. In addition, our
observation provides evidence of compatibility between the simulation
hypothesis and emotion theory, specifically as our concept of pleasure
is based on cognitive emotion theory.
Simulation hypothesis, emotional feedback, reinforcement learning.
Back to top
Paper #22
Having it both ways
the impact of fear on eating and fleeing in virtual flocking animals
Ruth Aylett
Professor of Computer Science
Mathematics and Computer Science, Heriot-Watt University
Edinburgh EH14 4AS, UK
The paper investigates the role of an affective
system as part of an
ethologically-inspired action-selection mechanism for virtual animals
in a
3D interactive graphics environment. It discusses the integration
of emotion
with flocking and grazing behaviour and a mechanism for communicating
emotion between animals. We develop a metric for analyzing the collective
behaviour of the animals and its complexity. We show that emotion
reduces
the complexity of behaviour and thus mediates between individual
and
collective behaviour.
Artificial life, autonomous agents, cognitive modeling, lifelike
characters, virtual reality
Back to top
Paper #23
On Compromise Strategies for Action Selection
with Proscriptive Goals
Frederick "Ric" Crabbe, USNA,
Devising
a solution to the action selection problem rests on a definition
of which actions are preferable over others. Among many properties
suggested for action selection mechanisms, one prominent one is
the ability to select compromise actions, i.e. actions that are not the
best to satisfy any active goal in isolation, but rather compromise
between the multiple goals. This paper performs an analysis
of compromise actions in situations where the agent has one proscriptive
goal. It concludes that optimal compromise behavior looks
quite different from what was expected, and, while optimal compromise
actions are beneficial to an agent, the benefit is often small
compared to greedy algorithms. It goes on to suggest that much of the
discussion about compromise behavior is the result of an equivocation
on its definition, and it concludes with a new compromise behavior
hypothesis.
Action
Selection; Compromise behavior; Optimal Behavior.
Back to top
Paper #24
Visual Communication
and Social Structure – The Group Predation of Lions
Alwyn Barry and Hugo Dalrymple-Smith
University of Bath
Department of Computer Science
Claverton Down, Bath, UK.
[Creel, 1997] in a study of african hunting dogs suggested that, where the maximisation of net
energy gain from hunting requires cooperation, cooperative hunting becomes an important part of the sociality of the hunting dogs.
When considering Lion cooperative hunting [Scheel and Packer, 1991] suggested, in contrast, that Lions do not
form groups to increase the intake of food for the
group but for the individual, implying that cooperative
behaviour in hunting has very little impact on the formation of prides. In using simulation to investigate
the role of visual location in the group hunting
behaviour of Lions it is shown that a minimal communication simulation can be derived if the dominance
of pride members is taken into account. We conclude that agreed dominance permits the reduction
of visual cues required to coordinate complex cooperative simulated behaviour.
Back to top
Paper #25:
Action selection in a macroscopic model
of the brainstem
reticular formation
Mark
Humphries, Kevin Gurney and Tony Prescott
University of Sheffield
The behavioral repertoire
of decerebrate and neonatal animals suggests that a relatively
self-contained neural substrate of action selection
may exist in the brainstem. Here, we develop
the hypothesis that the principal component of the substrate
is the medial ponto-medullary reticular formation.
Our quantitative structural model of this region,
which proposes a macroscopic organisation at
the level of inter-connected neural clusters,
is extended to incorporate sensory input. Evidence is reviewed
in support of the proposal that
both input and output configurations of this region
follow this organisation.
To investigate how this biologically constrained
model may be configured to support
action selection, a computational neural-population
model of the medial reticular
formation is outlined, and alternate configurations
are assessed in simulation. We conclude that the configuration
which most effectively supports action
selection is likely to be one which represents compatible
sub-actions at the cluster level; thus,
co-activation of a set of these clusters would lead to
the co-ordinated behavioral response observed in the animal.
Reticular formation, macroscopic
models, decerebrate
Back to top
Paper #26
Collective Action Selection in Social
Insect Colonies
James A. R. Marshall
Department of Computer Science
University of Bristol
Action selection is a problem that is faced not
only by individual organisms, but also by collectives
of organisms. Colonies of social insects
are highly integrated units that frequently
perform optimal collective action selection in
a decentralised manner. Social insect colonies
provide very accessible model systems for
studying the mechanisms underlying such action
selection processes. This paper discusses models of two action selection mechanisms in
insect colonies, and speculates as to the potential
for comparing action selection in such
colonies to action selection in individuals.
Collective action selection, individual action
selection, Temnothorax albipennis, Apis mellifera
Back to top
Paper #27
The basal ganglia as a cognitive response
mechanism: simulating normal and schizophrenic performance in the Stroop task
Tom Stafford and Kevin Gurney
Department of Psychology
University of Sheffield
Western Bank, Sheffield
S10 2TP, UK
This paper builds on our existing, biologically constrained model
of the
basal ganglia, originally constructed under the premise that these
subcortical structures perform action selection. Here, we show how
this same model, when used in conjunction with a connectionist model
of processing in the Stroop task, can also provide an account of normal
and pathological (schizophrenic) human reaction time performance. Our
model accounts for a wide variety of phenomenon and is consistent with
the most recent data on the physiology of schizophrenia. This work validates
modelling the basal ganglia as the vertebrate solution to the action
selection problem. Conversely, it demonstrates the importance of
issues of action selection and choice of biologically plausible response
mechanisms to performance on cognitive tasks.
Cognitive, response selection, stroop, schizophrenia, basal ganglia
Back to top
Paper #28
It's About Time: representing
time in cognitive models
Neil Madden and Brian Logan
We outline a cognitive architecture
which can be used to model a variety of perceptual phenomena.
The architecture is based on processes
operating on collections of time-limited
buffers in a parallel model of cognition
and draws on aspects of the Multiple Drafts theory \cite{Dennett/Kinsbourne:92a}.
We briefly describe the architecture and show how it can be used
to model two relevant experiments from the literature: colour phi
\cite{Kolers/vonGrunau:76a}, and the cutaneous ``rabbit'' \cite{Geldard/Sherrick:72a}.
Cognitive modelling, philosophical foundations,
AI architectures.
Back to top
Paper #29
Modelling Primate
Task Learning Requires Bad Machine Learning
Joanna J. Bryson and Jonathan C. S. Leong
University of Bath
Artificial models of natural Intelligence
Bath, BA2 7AY United Kingdom
Jonathan C. S. Leong
Harvard University
Primate Cognitive Neuroscience
Cambridge MA 02138, USA
jleong@fas.harvard.edu
We present a model of transitive inference which
is able to account for the performance of monkeys
and children on three-item transitive inference
tasks. We do this using a modular multi-layer neural
architecture which does not integrate error across layers. This system gets trapped in local minimae,
and in so doing, generates errors much like those seenin monkeys and children.
Back to top
Paper #30
Biorealistic Simulation of Baboon Foraging
using Agent-Based Modelling
Brian Logan
School of Computer Science & IT, University
of Nottingham
Phone: +44 115 846 6509, Fax: +44 115 951 4254
We present an agent-based model of the key activities
of a troop of chacma baboons (Papio hamadryas ursinus) based on data collected
at the De Hoop Nature Reserve in South Africa. We analyse the
predictions of the model in terms of how well it was able to duplicate the observed
activity patterns of the animals and the relationship between the parameters
that control the agent's decision procedure and the model's predictions.
The model predicts reasonable yearly average values for energy
intake, time spent socialising and resting, and habitat utilisation, but
is unable to account for month by month variation in the field data. However
even at the current stage of model development we are able to show
that, across a wide range of decision parameter values, the baboons are easily
able to achieve energetic and social time requirements. This
suggests that these particular animals are strongly influenced by other factors
such as predation risk or thermal load in deciding their activity
patterns.
Group action selection, primates
Back to top
Paper #31
Contracting model
of the basal ganglia
Benoˆýt Girard(1), Nicolas Tabareau(1) and Jean-Jacques
Slotine(2) and Alain Berthoz(1)
1.
Laboratoire de Physiologie de la Perception et de l’Action, CNRS
- Coll`ege de France 11 place Marcelin Berthelot, 75231 Paris Cedex
05, France.
2.
Non Linear System Laboratory, Massachusetts Institute of Technology,
Cambridge, Massachusetts, 02139, USA
It is thought that one role of the basal ganglia
is to constitute the neural substrate of action selection.
We propose here a modification of the action selection
model of the basal ganglia of (Gurney et al.,
2001a,b) so as to improve its dynamical features.
The dynamic behaviour of this new model is assessed
by using the theoretical tool of contraction
analysis. We simulate the model in the standard
test defined in (Gurney et al., 2001b) and also show
that it performs perfect selection when presented a
thousand successive random entries. From a biomimetical
point of view, our model takes into account a
usually neglected projection from GPe to the striatum,
which enhances its efficiency.
Contraction analysis, action selection,
basal ganglia, computational model
Back to top
Paper #32
Tolerance and Sexual
Attraction in Despotic Societies: A Replication and Analysis of
Hemelrik (2002)
Hagen Lehmann and JingJing Wang and Joanna J. Bryson
University of Bath
Artificial models of natural Intelligence
Bath, BA2 7AY United Kingdom
Most primate societies are characterised by hierarchical
structures with more or less despotic value.
Males are usually dominant over females, but in
periods of sexual attraction (during females period
of tumescence) male ‘tolerance’ towards females
rises. (5) shows in a model that this ‘tolerance’
is created as a side effect due to the rise of
female dominance during periods of sexual attraction.
This rise is the consequence of the more frequent
approaches of males towards females during
these periods. In Hemelrijk’s model the males
gain no benefit from ‘tolerating’ females and they
only do so at high aggression levels as a kind of ‘respectful
timidity’, because some of the females have become
dominant over them.
In this paper we discuss and replicate results
of Hemelrijk’s study about the effects of sexual
attraction and intensity of aggression between sexes
in despotic primate societies. We have used the
description of her individual-based model called
‘DomWorld’ to create our own version of this
model, which simulates an artificial primate society.
The replication has been successful, however
in implementing it we have found that some results
are highly reliant on aspects of the model that
are not well supported by the current primate
literature.
We discuss which of her results we find most
convincing, and relate this to the cognitivist debate
with special focus on macaque and chimpanzee societies.
Back to top
Paper #33
Toward an Executive without a Homunculus: Computational Models of the
Prefrontal Cortex/Basal Ganglia System.
Randall O'Reilly
The prefrontal cortex has long been thought to subserve both working memory
(the holding of information online for processing) and ``executive''
functions (deciding how to manipulate working memory and perform processing).
Although many computational models of working memory have been developed, the
mechanistic basis of executive function remains elusive, often amounting to a
homunculus. I present an attempt to deconstruct this homunculus through
powerful learning mechanisms that allow a computational model of the
prefrontal cortex to control both itself and other brain areas in a
strategic, task-appropriate manner. These learning mechanisms are based on
subcortical structures in the midbrain, basal ganglia and amygdala, which
together form an actor/critic architecture. The model's performance compares
favorably with standard backpropagation-based temporal learning mechanisms on
the challenging 1-2-AX working memory task, and other benchmark working
memory and cognitive control tasks. It also makes a number of testable
predictions about the contributions of the basal ganglia and prefrontal
cortex in various behavioral tasks, several of which have been tested and
confirmed.
Back to top
Paper #34
Neurocomputational modeling of human decision-making
Marius Usher
Decision-making is one of the most common and, at the same time,
open-ended and effortful human activities. One source of its difficulty
resides with the need to evaluate alternatives whose `attractiveness'
varies on several incommensurable dimensions. Experimental work in human
decision-making has also revealed a series of intriguing behavioral
patterns that indicate deviations from normative economic theories and
which raise a challenge for the development of a theory of human
performance. Here I will review some of these patterns, such as
loss-aversion and preference reversals under a series of conditions (time
constraints, the introduction of contextual information in the choice set,
etc). I will then discuss and contrast a number of neurocomputational
theories that have recently been proposed to account for these patterns
and to explain the cognitive processes that mediate choice-RT and human
decision-making.
Back to top
Paper #35
Endogenous Political Parties
Michael Laver, New York University
Michel Schilperoord, Erasmus University, Rotterdam
The spatial model of party competition is one of the dominant paradigms
of contemporary political science. Virtually all spatial models of
party competition are essentially static: most key model parameters,
including the identity of all parties and rules of interaction between
them, are set exogenously; the essential solution concept deployed is
some form of static equilibrium. However, recent progress has been made
with agent-based models that treat party competition as an evolving
complex system that may never reach a steady state, (Kollman, Miller
and Page 1992; Kollman, Miller and Page 1998; De Marchi 1999; De Marchi
2003; Kollman, Miller and Page 2003; Laver 2005). The central purpose
of this paper is to extend the agent-based model of party competition
proposed in Laver (2005) to encompass the birth and death of political
parties and thereby make the identity of parties in the system an
output of, rather than an analyst-specified input to, the model. This
is done by modeling party birth as an endogenous change of agent type
from citizen to party leader. In order to do this it becomes necessary
to model the "memories" of citizens in the system, an issue that has
not previously been addressed in agent-based models of party
competition, which have hitherto assumed goldfish memories. The birth
and death of parties transforms into a dynamic system even an
environment where all agents have otherwise non-responsive adaptive
behaviors. Substantively, the original purpose of this modeling
exercise was to investigate how key system parameters condition the
number and identity of political parties in a given system. An
unintended but valuable spin-off has been that we are now able to
characterize the overall social welfare of the set of citizens, taken
as a whole, as a function of party system parameters.
Back to top