Cognitive Computation: A Case Study in Cognitive Control of Autonomous Systems and Some Future Directions - PowerPoint PPT Presentation

Loading...

PPT – Cognitive Computation: A Case Study in Cognitive Control of Autonomous Systems and Some Future Directions PowerPoint presentation | free to download - id: 475da4-ODQxM



Loading


The Adobe Flash plugin is needed to view this content

Get the plugin now

View by Category
About This Presentation
Title:

Cognitive Computation: A Case Study in Cognitive Control of Autonomous Systems and Some Future Directions

Description:

Cognitive Computation: A Case Study in Cognitive Control of Autonomous Systems and Some Future Directions Professor Amir Hussain, Dr Andrew Abel – PowerPoint PPT presentation

Number of Views:181
Avg rating:3.0/5.0
Slides: 94
Provided by: Erfu
Learn more at: http://www.cs.stir.ac.uk
Category:

less

Write a Comment
User Comments (0)
Transcript and Presenter's Notes

Title: Cognitive Computation: A Case Study in Cognitive Control of Autonomous Systems and Some Future Directions


1
Cognitive Computation A Case Study in Cognitive
Control of Autonomous Systems and Some Future
Directions
  • Professor Amir Hussain, Dr Andrew Abel
  • 1 Division of Computing Science and Mathematics
  • University of Stirling, Scotland
  • Work reported here is part of an ongoing UK EPSRC
    funded project, with Dr Erfu Yang1 (RF) Prof
    Kevin Gurney2 (CI)
  • 2Adaptive Behaviours Research Group (ABRG)
  • Department of Psychology University of
    Sheffield, UK

The International Joint Conference on Neural
Networks (IJCNN) Dallas, Texas, August 4-9, 2013
2
Introduction
  • Why Cognitive Computation?
  • Why Cognitive Machines?
  • Taylors Proposal on Cognitive Machines
  • Cognitively Inspired Control of Autonomous
    Systems
  • Towards a more generalised cognitive framework

Cognitive Signal Image and Control Processing
Research (COSIPRA) Laboratory
3
Introduction
  • Cognitive computation
  • an emerging discipline linking together
    neurobiology, cognitive psychology and artificial
    intelligence
  • Springers journal Cognitive Computation
    publishing biologically inspired theoretical,
    computational, experimental and integrative
    accounts of all aspects of natural and artificial
    cognitive systems.
  • Professor John Taylor
  • founding Advisory Board Chair of Cognitive
    Computation
  • proposed on how to create a cognitive machine
    equipped with multi-modal cognitive capabilities.
  • This keynote
  • first presents a novel modular cognitive control
    framework for autonomous systems - potentially
    realizes the required cognitive action-selection
    and learning capabilities in Professor Taylor's
    envisaged cognitive machine.
  • Possible future avenues for improving this work
    in a cognitively inspired manner

Cognitive Signal Image and Control Processing
Research (COSIPRA) Laboratory
4
Why Cognitive Computation?
  • Promote a more comprehensive and unified
    understanding of diverse topics
  • perception, action, and attention
  • learning and memory
  • decision making and reasoning
  • language processing and communication
  • problem solving and consciousness aspects of
    cognition.
  • Industry, commerce, robotics and many other
    areas are increasingly calling for the creation
    of cognitive machines, with cognitive powers
    similar to those of ourselves
  • are able to think for themselves
  • reach decisions on actions in a variety of ways
  • are flexible, adaptive and able to learn from
    both their own previous experience and that of
    others around them

Cognitive Signal Image and Control Processing
Research (COSIPRA) Laboratory
5
Why Cognitive Machines?
  • A multi-disciplinary research challenge
  • Understanding our own cognitive powers
  • how they are created and fostered
  • how they can go wrong due to brain malfunction
  • the modelling of the cognitive brain is an
    important step in developing such understanding.
  • Creating autonomous robots and vehicles able to
    think and act cognitively and ethically
  • support us in our daily lives.

Cognitive Signal Image and Control Processing
Research (COSIPRA) Laboratory
6
Taylors Proposal on Cognitive Machines
  • It was published at J.G. Taylor, Cognitive
    computation, Cogn. Comput, vol.1, pp.416
    (2009).
  • Based on ideas published in many places
  • Taylor raised a number of very interesting points
    in his attempts to construct an artificial being
    empowered with its own cognitive powers
  • a range of key questions relevant to the creation
    of such a machine
  • made detailed and methodical attempts to answer
    these questions
  • providing convincing evidence from national and
    international research projects he had led over
    the years.
  • Taylors proposal is one of very few attempts to
    construct a global brain theory of cognition and
    consciousness.
  • It is based on a unique multi-modal approach
    that takes into consideration vision and
    attention, motor action, language and emotion.
  • Conventional studies in cognition and
    consciousness have mostly focussed on single
    modalities such as vision.

Cognitive Signal Image and Control Processing
Research (COSIPRA) Laboratory
7
Taylors Proposal on Cognitive Machines
  • Taylor asked a number of questions
  • What is human cognition in general, and how can
    it be modelled?
  • What are the powers of animal cognition, and how
    can they be modelled?
  • How important is language in achieving a
    cognitive machine, and how might it be developed
    in such a machine?
  • What are the benchmark problems that should be
    able to be solved by a cognitive machine?
  • Does a cognitive machine have to be built in
    hardware?
  • How can hybridisation help in developing truly
    cognitive machines
  • Is consciousness crucial?
  • How are the internal mental states of others to
    be discerned?
  • Discussed notion of attention control

Cognitive Signal Image and Control Processing
Research (COSIPRA) Laboratory
8
Taylors Proposal on Cognitive Machines
  • This approach to attention control relevant to
    our interests
  • Will link to a case study that uses this as a
    basis for a new approach to autonomous vehicle
    control
  • Initially focus on control and decision making
  • Ongoing work!

Cognitive Signal Image and Control Processing
Research (COSIPRA) Laboratory
9
Cognitive Control of Autonomous Systems
  • A Case Study

10
Two problem domains
  • Planetary rovers (SciSys)
  • Smart cars (Google)

11
Challenges in each domain
  • Urban driving in smart cars
  • constantly changing trajectories
  • moderated speed in urban areas
  • sentinel awareness of high pedestrian density
  • Planetary rovers
  • real-time trajectory planning for feasible path
    to follow on
  • Autonomous navigation
  • Intelligent motion control with most optimal
    controller
  • Active and smart obstacle avoidance
  • cognitive awareness of complex environments

Cognitive Signal Image and Control Processing
Research (COSIPRA) Laboratory
12
The problem we tacklefrom partially specified
trajectories to cognitive control
X(t0)
X(t1)
X(t2)
Path following with error correction Take account
of obstacles and challenges
13
The problem we tacklefrom partially specified
trajectories to cognitive control
Vehicle with given dynamics and kinematics Drives
along P(t)
X(t0)
X(t1)
X(t2)
14
Multiple controller methods
  • Historically
  • Hard switching
  • One controller selected at any one time
  • Issue is bumpiness when switching between
    controllers
  • Our goal
  • Bumpless control
  • Soft switching
  • Select a subset of all controllers
  • Mix controller decisions together
  • Smoother output

Cognitive Signal Image and Control Processing
Research (COSIPRA) Laboratory
15
Existing hard switching control
supervisor
disturbance/ noise
switching signal
s
measured output
w
reference input
controller 1
s
e(t)
r(t)
y
bank of candidate controllers
Plant Model

u
_
controller n
control signal
  • Key ideas
  • Build a bank of alternative controllers
  • Switch among them online based on switching
    condition

16
Compare with the problem of action selection in
animals
Fight, flight or feeding, but not do nothing
  • The animal solution is centred on a set of
    sub-cortical brain nuclei the basal ganglia,
    which act as a central switch or selector
  • Can we leverage the biological solutions for use
    in AVC?

Basal ganglia in brain
Cognitive Signal Image and Control Processing
Research (COSIPRA) Laboratory
17
The biology Disinhibition gating and action
channels(compare with modular control)
Predisposing conditions
Motor resources
Cognitive Signal Image and Control Processing
Research (COSIPRA) Laboratory
18
Modular control Challenges
  • Meeting multiple performance criteria
  • Stability
  • Convergence
  • Tackling problems of chattering
  • Anti-windup and Bumpless switching
  • Real-time operation

Cognitive Signal Image and Control Processing
Research (COSIPRA) Laboratory
19
Three-stage modular framework a bio-inspired
approach
Cognitive Signal Image and Control Processing
Research (COSIPRA) Laboratory
20
Using the biomimetic BG model in a control
environment
  • 4-wheel rover Kinematics-based motion control
    and planning

21
Three-stage modular framework case study
Actual trajectory
Cognitive Signal Image and Control Processing
Research (COSIPRA) Laboratory
22
Kinematics-based motion control and planning
  • The motion control of autonomous vehicles is
    mostly based on the vehicles kinematics model
  • Usually assumed that the vehicles internal
    dynamics can immediately satisfy the
    velocity/steering angle requests from the
    kinematics-based motion control
  • This study
  • BG-based kinematic motion controllers are used
    for motion planning and control
  • Perfect dynamics assumed

Cognitive Signal Image Processing and Control
Systems Research (COSIPRA) Laboratory
23
Kinematics-based motion control and planning
Feedback linearisation
actual trajectory
Kinematics to path
Two trajectory Components (input from motion
planner)
Controllers are all Pole placement-based
Cognitive Signal Image Processing and Control
Systems Research (COSIPRA) Laboratory
24
Action surface for fuzzy salience model
Cognitive Signal Image and Control Processing
Research (COSIPRA) Laboratory
25
Simulation Results
  1. Circular Trajectory Tracking Control

(b) x - y trajectory comparison for BG-based
switching and a single feedback linearization
motion controller under noise
(a) States in the circular tracking with BG-based
switching and a single feedback linearization
motion controller under noises
Cognitive Signal Image Processing and Control
Systems Research (COSIPRA) Laboratory
Cognitive Signal Image and Control Processing
Research (COSIPRA) Laboratory
26
C. General Path Tracking double lane change and
roundabout
x-y trajectory under BG-based switching and a
single feedback linearization motion controller
under noises
Cognitive Signal Image Processing and Control
Systems Research (COSIPRA) Laboratory
Cognitive Signal Image and Control Processing
Research (COSIPRA) Laboratory
27
Using the biomimetic BG model in a control
environment
  • 4-wheel rover B-Spline path planning and
    three-stage motion control with integrated
    kinematics and dynamics

28
Smooth path planning with B-splines
  • The dimension of the knot vector 24
  • The number of control points 18
  • The degree of splines 5

Control points and smooth path planned with
B-spline method
Cognitive Signal Image Processing and Control
Systems Research (COSIPRA) Laboratory
Cognitive Signal Image and Control Processing
Research (COSIPRA) Laboratory
29
General Path Tracking double lane change and
roundabout
Comparison of BG-based soft switching control and
single-fixed controller with noises
Cognitive Signal Image Processing and Control
Systems Research (COSIPRA) Laboratory
Cognitive Signal Image and Control Processing
Research (COSIPRA) Laboratory
30
Comparison of Control Performance (MSE Mean
Squared Error)
Performance BG without noise BG with noise Single without noise Single with noise
MSE in x 0.0044 0.0046 0.0565 0.0652
MSE in y 0.0000016832 0.00090293 0.000014852 0.0020
MSE in x-y 0.0031 0.0033 0.04 0.0461
Cognitive Signal Image Processing and Control
Systems Research (COSIPRA) Laboratory
Cognitive Signal Image and Control Processing
Research (COSIPRA) Laboratory
31
Summary
  • BG-based controller selection is bumpless
    soft-switching because it combines outputs of
    multiple controllers
  • We have some evidence that this also helps avoid
    windup chattering
  • BG will allow adaptive control by varying
    internal parameters which are now better
    understood from our neurobiological models
  • Based on model of biological decision making
  • Attention switching using salience
  • Ongoing work

Cognitive Signal Image Processing and Control
Systems Research (COSIPRA) Laboratory
32
Autonomous Control Specific Future Work
  • Test against traditional switched controller
    designs with same controllers
  • Adaptive online operations
  • learn salience weights to BG controller
  • Dynamic allocation of controllers
  • Use of more realistic models
  • Real experimental test beds

Cognitive Signal Image Processing and Control
Systems Research (COSIPRA) Laboratory
33
Cognitive Future Work
  • Incorporating vision
  • Better able to react to world
  • Use of multiple modalities
  • Dual process control.
  • Automatic behaviour mode
  • Process known differently from unknown
  • Learning over time, becomes automatic
  • Mimics processing in the brain

Cognitive Signal Image Processing and Control
Systems Research (COSIPRA) Laboratory
34
Cognitive Computation
  • towards a multimodal framework

Cognitive Signal Image Processing and Control
Systems Research (COSIPRA) Laboratory
35
More Cognitive Computation?
  • This is a specific case study
  • Inspired by work of John Taylor
  • Cognitive Computation is very wide ranging field
    of research
  • Can be applied in many different contexts
  • Means different things to different people
  • Presentation tomorrow
  • Discuss cognitive computation in more depth
  • Application in more fields
  • Want to consider a more general cognitive
    framework

Cognitive Signal Image Processing and Control
Systems Research (COSIPRA) Laboratory
36
Sentic Computing
  • Sentiment Analysis
  • Common sense computing
  • Read emotion and tone from text
  • Traditional approaches inadequate
  • Machine Learning
  • Keyword counting
  • May identify topic, but not sentiment
  • Concept based approach
  • Can assign emotions to concepts
  • Relate similar concepts together

Cognitive Signal Image Processing and Control
Systems Research (COSIPRA) Laboratory
37
AffectNet Graph
Cognitive Signal Image Processing and Control
Systems Research (COSIPRA) Laboratory
38
AffectiveSpace
  • E. Cambria and A. Hussain. Sentic Computing
    Techniques, Tools, and Applications. Dordrecht,
    Netherlands Springer, ISBN 978-94-007-5069-2
    (2012)

Cognitive Signal Image Processing and Control
Systems Research (COSIPRA) Laboratory
39
Sentic Computing
  • Sentic Activation
  • Consider conscious and unconscious level
    processing
  • The two interact
  • Can be used for sentiment analysis
  • Emotion detection

Cognitive Signal Image Processing and Control
Systems Research (COSIPRA) Laboratory
40
Multimodal Speech Processing
  • Traditional hearing aids focus on single modality
  • This is not the whole story!
  • Perception, attention switching
  • Multimodality
  • McGurk effect
  • Lip reading used in noisy environments
  • More extensively by those with hearing problems
  • Visual information used, but only when
    appropriate
  • Conscious and unconscious processing
  • Speech often works on prediction

Cognitive Signal Image Processing and Control
Systems Research (COSIPRA) Laboratory
41
Multimodal Speech Processing
  • A different direction for listening devices and
    hearing aids
  • Consider how people actually hear
  • Lip reading as part of speech filtering
  • Cognitively inspired nuanced use of visual
    information

Cognitive Signal Image Processing and Control
Systems Research (COSIPRA) Laboratory
42
General Cognitive Framework
  • Taylor discussed the creation of a cognitive
    being
  • Language
  • Consciousness
  • Decision making
  • Memory
  • Emotional coding
  • Aim is to consider a more general purpose
    approach
  • Basal Ganglia inspired decision making
  • Concept based emotion analysis
  • Multimodal speech interpretation capabilities
  • Dual level processing
  • Can they be combined into a multimodal framework?

Cognitive Signal Image Processing and Control
Systems Research (COSIPRA) Laboratory
43
General Cognitive Framework
  • Multimodality
  • More environmentally aware
  • Additional sensors to feed into a vehicle control
    system
  • Vision, sound, weather conditions etc.
  • Communication
  • Communicate with those in the car and outside
  • Speech recognition and generation
  • Sentiment analysis from passengers
  • Able to learn and adapt to wishes of those in car
  • Adjust behaviour to suit conditions and emotions
  • Multimodal social and cognitive agents

Cognitive Signal Image Processing and Control
Systems Research (COSIPRA) Laboratory
44
Sentic Blending Scalable Multimodal Fusion for
the Continuous Interpretation of Semantics and
Sentics
  • A general and scalable methodology termed sentic
    blending, for interpreting the conceptual and
    affective information associated with natural
    language through different modalities
  • enables the continuous interpretation of
    semantics and sentics (i.e., the conceptual and
    affective information associated with natural
    language)
  • based on the integration of an affective
    common-sense knowledge base with any multimodal
    cognitive signal image and control processing
    module.
  • operates in a multidimensional space that enables
    the generation of a continuous stream
    characterizing users semantic and sentic
    progress over time - despite the outputs of
    the unimodal categorical modules having very
    different time-scales and output labels.
  • Uses decision fusion

Cognitive Signal Image Processing and Control
Systems Research (COSIPRA) Laboratory
45
A sample schema of continuous multimodal fusion
through sentic blending
Cognitive Signal Image and Control Processing
Research (COSIPRA) Laboratory
46
An application example SenticNet Engine
Ensemble streams obtained when applying sentic
blending to the SenticNet engine (left) and the
facial expression analyser (right), without
sentic kinematics filtering.
Cognitive Signal Image and Control Processing
Research (COSIPRA) Laboratory
47
An application example SenticNet Engine
Ensemble stream obtained when applying sentic
blending to the proposed conversation, with
(right) and without (left) using sentic
kinematics filtering.
Cognitive Signal Image and Control Processing
Research (COSIPRA) Laboratory
48
Performance Comparison
Confusion matrix obtained combining the five
classifiers. Success rates for neutral, joy, and
surprise are very high, but disgust, anger, and
fear tend to be confused
Confusion matrix obtained after human assessment.
Success ratios considerably increase, meaning
that the adopted classification strategy is
consistent with human classification.
49
General Cognitive Framework
  • Considers the emotional states of others
  • Considers aspects of human cognition
  • Considers the issue of language
  • Considers benchmark problems
  • Convincing communication
  • Could be extended to include vehicle and language
    control
  • Driving, extremely challenging problem
  • Dual level processing
  • Cognitively inspired use of different modalities
  • Dual layer processing is unifying

Cognitive Signal Image Processing and Control
Systems Research (COSIPRA) Laboratory
50
Acknowledgements
  • Everyone who helped to organise this conference!
  • All of the COSIPRA Lab
  • http//cosipra.cs.stir.ac.uk
  • Dr Erfu Yang, Prof Leslie Smith, Dr Erik Cambria

Cognitive Signal Image Processing and Control
Systems Research (COSIPRA) Laboratory
51
  • Thanks for listening!
  • Questions?

Cognitive Signal Image Processing and Control
Systems Research (COSIPRA) Laboratory
52
Appendix
53
Two Modes of Biological Action Selection
Automatic/Habitual and Controlled/Executive
Processing - I
  • In psychological literature, modes of behavioural
    control refer to automatic (or habitual)
    controlled (or executive) processing respectively
    with their joint use constituting a dual-process
    theory of behaviour
  • Controlled processing is under the subjects
    direct and active control, is slow, and requires
    serial attention to component stimuli or
    sub-tasks. In contrast, automatic control is less
    effortful, less prone to interference from
    simultaneous tasks, is driven largely by the
    current stimulus and does not necessarily give
    rise to conscious awareness
  • Dual-process theory also supposes a dynamic
    transfer of control under learning.
  • The development of automatic processing has close
    similarities with the notion of stimulus-response
    (S-R) learning, or habit learning.
  • Controlled processing may be likened to
    goal-directed behaviour in animals.

Cognitive Signal Image and Control Processing
Research (COSIPRA) Laboratory
54
Two Modes of Biological Action Selection
Automatic/Habitual and Controlled/Executive
Processing - II
  • Habits are supported in closed-loop circuits
    through BG associated with sensorimotor cortical
    areas.
  • The pre-frontal-cortex (PFC) serves as an
    executive' or supervisory role in enabling
    controlled processing. PFC also forms loops
    through BG. The supervisory' PFC works to
    modulate or bias the action selection of the
    automatic (sensorimotor) processing system.
  • Controlled processing dominates in the early
    acquisition of new skills which subsequently,
    when well-practiced, are carried out using
    automatic processing.
  • As in dual-process theory, it is supposed that
    goal-directed (non-habitual) behaviours governed
    by PFC can transfer into habits in sensorimotor
    loops by learning therein under the influence of
    the PFC loops

Cognitive Signal Image and Control Processing
Research (COSIPRA) Laboratory
55
Kinematics Vehicle model for Motion Control and
Planning
Yang, Hussain, and Gurney. (BICS 2013) to appear.
imposes the physical constraint, - the steering
angle delta is contrained within a desired
(state) range (to enable a smooth time
invariant control solution)
If the steering angle is selected as one control
input, then the kinematics model can be further
simplified as
56
Advanced Motion Controller Method I/O Feedback
Linearization Controller Design Process
Erfu Yang, Amir Hussain, and Kevin Gurney. A
basal ganglia inspired soft switching approach to
the  motion control of a car-like autonomous
vehicle. The 2013 International Conference on
Brain Inspired Cognitive Systems (BICS 2013),June
9-11, 2013, Beijing, China, to appear.
57
Fuzzy logic rules for BG-Based soft switching
motion control
Cognitive Signal Image and Control Processing
Research (COSIPRA) Laboratory
58
Path/Trajectory Planning Consider two
sixth-order polynomials of time t and their
derivatives
59
(No Transcript)
60
Generic Solution
61
Dynamics vehicle model used (for car-like rover)
Eric N Moret. Dynamic Modeling and Control of a
Car-Like Robot. Thesis, Virginia Polytechnic
Institute and State University,2003.
Cognitive Signal Image and Control Processing
Research (COSIPRA) Laboratory
62
Gain-Scheduling vs BG control
  • Another important idea formed in this project
    thus far is to utilize the reference signal as a
    priori knowledge of the control system under
    consideration to aid the realization of
    automatic (habitual) mode behaviour.
  • This shares some similarity with traditional
    gain-scheduling solution in which a family of
    controllers such as PI or PID related to the
    control reference signal and desired output are
    designed (Zhao et al, ).
  • An engine control model for autonomous vehicle
    has been employed initially to illustrate this
    traditional gain-scheduling approach.

Cognitive Signal Image and Control Processing
Research Laboratory
63
Gain-Scheduling vs. Reference-based habits?
  • 11 PI controllers are demonstrated.
  • So, the action (controller) selection in the
    automatic mode can be realized by mapping the
    reference signal (desired engine speed in the
    case) to the controllers parameters (gains).
  • In our proposed BG-based soft switching approach,
    this action selection can be realised in a more
    natural way, which will be demonstrated further
    in the vehicles cognitive cruise control - NEXT

Cognitive Signal Image and Control Processing
Research Laboratory
64
Appendix B
  • Sentic blending

65
Sentic Blending Scalable Multimodal Fusion for
the Continuous Interpretation of Semantics and
Sentics
  • Aimed at extending the modular cognitive
    framework to incorporate additional modalities
  • by integrating vision, language and emotion
  • for enabling multi-modal social cognitive and
    affective behavioural capabilities in autonomous
    agents.
  • A general and scalable methodology termed sentic
    blending, for interpreting the conceptual and
    affective information associated with natural
    language through different modalities
  • enables the continuous interpretation of
    semantics and sentics (i.e., the conceptual and
    affective information associated with natural
    language)
  • based on the integration of an affective
    common-sense knowledge base with any multimodal
    cognitive signal image and control processing
    module.
  • operates in a multidimensional space that enables
    the generation of a continuous stream
    characterizing users semantic and sentic
    progress over time - despite the outputs of
    the unimodal categorical modules having very
    different time-scales and output labels.

Cognitive Signal Image and Control Processing
Research (COSIPRA) Laboratory
66
A sample schema of continuous multimodal fusion
through sentic blending
Cognitive Signal Image and Control Processing
Research (COSIPRA) Laboratory
67
An application example SenticNet Engine
Ensemble streams obtained when applying sentic
blending to the SenticNet engine (left) and the
facial expression analyser (right), without
sentic kinematics filtering.
Cognitive Signal Image and Control Processing
Research (COSIPRA) Laboratory
68
An application example SenticNet Engine
Ensemble stream obtained when applying sentic
blending to the proposed conversation, with
(right) and without (left) using sentic
kinematics filtering.
Cognitive Signal Image and Control Processing
Research (COSIPRA) Laboratory
69
Performance Comparison
Confusion matrix obtained combining the five
classifiers. Success rates for neutral, joy, and
surprise are very high, but disgust, anger, and
fear tend to be confused
Confusion matrix obtained after human assessment.
Success ratios considerably increase, meaning
that the adopted classification strategy is
consistent with human classification.
70
Appendix C
  • Attention control

71
Taylors Attention Control
Goal Module
ATTN Signal Creator
Input Module
  • Ballistic Attention Control System

Goals
Attention Controller
Cortex
ATTN Signal Creator
Input Module
WM cd
Objects/Features
Monitor
ATTN Copy Module
Buffer Memory
Wm input
  • The corollary discharge of attention model
    (CODAM) for consciousness
  • Attention copy of Attention Control

72
Multiple controller methods
  • One promising approach to AVC is to break the
    task into sub-tasks, each valid over a restricted
    range of conditions, and to switch between them
    when required, based on sensory and internally
    generated signals.
  • Historically achieved using several approaches
    such as
  • PIDGain scheduling (Ahmad 09)
  • Sliding mode control
  • Dynamic feedback linearisation (Oriolo 02
    Kulkarn,NASA JPL )
  • Fuzzy logicPIDmultiple models (Iagnemma 99,
    MIT Narendra, Yale Hussain Gurney et al.
    08,09, Stirling)
  • Neural approaches (Shumeet 96, Kawato Wolpert,
    2001)
  • Decision-theoretic control (Zilberstei,02)

Cognitive Signal Image and Control Processing
Research (COSIPRA) Laboratory
73
Appendix D - A biological interlude
  • Basal ganglia and action selection

74
BG Functional Model
S1 S2 S3
Feedforward Off-centre, on-surround network
Z1 Z2 Z3
(Gurney et al. 2001)
Cognitive Signal Image and Control Processing
Research (COSIPRA) Laboratory
75
Vector inputs effective salience
  • Effective salience s (scalar), with input vector
    x, and channel weight vector w, is given by s
    f(w, x)
  • s f(w, x) may be simple dot product or
    arbitrary nonlinear function

Cognitive Signal Image and Control Processing
Research (COSIPRA) Laboratory
76
Appendix E - Using the biomimetic BG model in a
control environment
  • 4-wheel rover Kinematics-based motion control
    and planning

77
Three-stage modular framework case study
Actual trajectory
Cognitive Signal Image and Control Processing
Research (COSIPRA) Laboratory
78
Kinematics-based motion control and planning
  • The motion control of autonomous vehicles is
    mostly based on the vehicles kinematics model
  • Usually assumed that the vehicles internal
    dynamics can immediately satisfy the
    velocity/steering angle requests from the
    kinematics-based motion control
  • This study
  • BG-based kinematic motion controllers are used
    for motion planning and control
  • Perfect dynamics assumed

Cognitive Signal Image Processing and Control
Systems Research (COSIPRA) Laboratory
79
Kinematics-based motion control and planning
Feedback linearisation
actual trajectory
Kinematics to path
Two trajectory Components (input from motion
planner)
Controllers are all Pole placement-based
Cognitive Signal Image Processing and Control
Systems Research (COSIPRA) Laboratory
80
  • Each controller has different parameters
  • One salience, one controller
  • 300 controllers
  • Sub tasks following path

81
  • Input signals (x,y) separated
  • Each input fed into all controllers
  • Each controller is different
  • Outputs a recommended action
  • Signal and error also fed into fuzzy logic
  • Determines salience,
  • urgency, based on error and reference
  • Apply to basal ganglia model
  • Selection strength of each controller
  • Gating function to normalise
  • Between 0 and 1
  • Gating function output applied to each controller
  • Acts as a weight, could be zero
  • Outputs summed
  • Recoupled to determine output
  • See BICS 2013 paper

82
Action surface for fuzzy salience model
Cognitive Signal Image and Control Processing
Research (COSIPRA) Laboratory
83
  • Each one represents a different salience output
  • Essentially, each one reacts differently

84
Simulation Results
  1. Circular Trajectory Tracking Control

(b) x - y trajectory comparison for BG-based
switching and a single feedback linearization
motion controller under noise
(a) States in the circular tracking with BG-based
switching and a single feedback linearization
motion controller under noises
Cognitive Signal Image Processing and Control
Systems Research (COSIPRA) Laboratory
Cognitive Signal Image and Control Processing
Research (COSIPRA) Laboratory
85
  • Currently only single controller
  • Testing against hard controller currently

86
B. Lane Change
(b) x - y trajectory comparison for BG-based
switching and a single feedback linearization
motion controller under noise
(a) States under BG-based switching and a single
feedback linearization motion controller under
noises
Cognitive Signal Image Processing and Control
Systems Research (COSIPRA) Laboratory
Cognitive Signal Image and Control Processing
Research (COSIPRA) Laboratory
87
C. General Path Tracking double lane change and
roundabout
x-y trajectory under BG-based switching and a
single feedback linearization motion controller
under noises
Cognitive Signal Image Processing and Control
Systems Research (COSIPRA) Laboratory
Cognitive Signal Image and Control Processing
Research (COSIPRA) Laboratory
88
Using the biomimetic BG model in a control
environment
  • 4-wheel rover B-Spline path planning and
    three-stage motion control with integrated
    kinematics and dynamics

89
  • B spline generates smoother path

90
Smooth path planning with B-splines
  • The dimension of the knot vector 24
  • The number of control points 18
  • The degree of splines 5

Control points and smooth path planned with
B-spline method
Cognitive Signal Image Processing and Control
Systems Research (COSIPRA) Laboratory
Cognitive Signal Image and Control Processing
Research (COSIPRA) Laboratory
91
  • What is spline?
  • What is the knot vector, control parameter,
    controlling?

92
General Path Tracking double lane change and
roundabout
Comparison of BG-based soft switching control and
single-fixed controller with noises
Cognitive Signal Image Processing and Control
Systems Research (COSIPRA) Laboratory
Cognitive Signal Image and Control Processing
Research (COSIPRA) Laboratory
93
Comparison of Control Performance (MSE Mean
Squared Error)
Performance BG without noise BG with noise Single without noise Single with noise
MSE in x 0.0044 0.0046 0.0565 0.0652
MSE in y 0.0000016832 0.00090293 0.000014852 0.0020
MSE in x-y 0.0031 0.0033 0.04 0.0461
Cognitive Signal Image Processing and Control
Systems Research (COSIPRA) Laboratory
Cognitive Signal Image and Control Processing
Research (COSIPRA) Laboratory
About PowerShow.com