Presentacin de PowerPoint

1
An Immune-based Multilayered Cognitive Model for
Autonomous Navigation
Diego A. Romero, M. Sc. Electrical and
Electronics Engineering Dept. Fernando Niño, Ph.
D. Computer Engineering Dept. National
University of Colombia
2

• GOAL
• Immune-based decision making model, architecture
  based on multilayer processing
• To test the model on a mobile agent that solves
  non-trivial tasks on a two-dimensional
  environment
• Autonomous navigation systems able to adapt to
  dynamic environments in order to achieve their
  goals.

3

• Outline
• Introduction
• Problem Definition
• Proposed Immune-based Cognitive Model
• Basic Navigation
• Object Identification
• Trajectory Optimization
• 4. Experiments
• 5. Conclusions and Future Work

4

• 2. Navigation Problem Definition
• Autonomous agent that moves on a (simulated) two
  dimensional rectangular environment
• The environment contains a set of geometrically
  different objects
• Same number of storage locations as objects
• Obstacles (walls) distributed around the
  environment

5
Environment
Objects
Mobile agent
Storage locations
Obstacles and environment limits
6
Mobile agent Circular geometry (similar to
Khepera, Khepera II, Microbot Kity)

• 12 ultrasónic sensors located on its perimeter
• Each ultrasonic sensor has a 30 angle scope
• 1 photo-sensor, to distinguish objects from
  obstacles

7

• Decision levels of the problem
• Environment exploration to find the objects to
  reallocate
• To sense the perímeter of the objects to capture
  their geometrical features
• 2. Object identification in order to allocate
  them in their corresponding storage locations
• 3. Computing the shortest trajectory to take the
  objects to the storage locations, once they have
  been recognized

8
3. Proposed Model
Multilayer Architecture (Rodney Brooks)

• 3 cognitive layers
• Basic Navigation primary behavior, low level of
  complexity
• Object Identification intermediate level of
  abstraction
• Trajectory Optimization high level of
  abstraction

9
1st cognitive layer Basic Navigation The mobile
agent needs to learn to move in the environment
ensuring its physicall integrity It has to learn
to avoid colliding with objects and
obstacles Constraints Sensors Perception
constraints Sensor noise Uncertainty on the
agents location real-time requirements
10
Navigation Neural Network
Training patterns
Inputs
Outputs
S1 move forward S2 turn S3 30º (right or left)
Inputs

• First four(4) neurons front sensorsreadings
  (collision free movements, fine alignment with
  the object contour)
• Other eight (8) neurons inputs from the other
  sensors

48 training patterns
11
Navigation behaviors learned by the NN
Search
General border tracking
Align
Degree of Complexity
Scanning not very sharp edges
Avoid Obstacles
Collision-free navigation
Explore
12
Scanning an objects geometry
13
Information used by the second cognitive layer
Detecting a change of direction
14

• 2nd cognitive layer Object Identification
• Once the agent has found an object, it must
  identify it and then transport it to one of the
  storage locations.
• The agent should learn to identify each object
  through the association between its sensed
  physical features and its respective storage
  location

15
Artificial Cognitive Immune System (ACIS) Based
on an artificial immune network (AiNet)
reinforcement learning credit assignment
algorithm Bucket Brigade
16
ACIS
17
ACIS inputs
Antigen encoding
18
3rd Cognitive Layer. Trajectory optimization
layer Once ACIS has activated an Ab with its
identification response, the learning process
expects a reinforcement to the action taken In
order to get a reinforcement the agent needs to
collect the object and transport it to the
storage location identified by ACIS
19
First four inputs frontal sensors Other two
inputs angle to the goal storage location

• First bit clockwise 30 turn and moving
  forward.
• Second bit moving forward without turning
• Third bit counterclockwise 30 turn and
  moving forward.

20
Subsumption Architecture Approach Layer
interaction First and second layer interaction
Once the mobile agent finds and senses an
object via its navigation mechanism, it
classifies it through the second cognitive
layer Second and third layer interaction When
the agent has learned to identify all the objects
via the interaction between first and second
layer, the cognitive model implemented in the
third layer utilizes the final classification to
generate an optimal trajectory. First cognitive
layer is trained off-line. NO on-line
learning The second and third layer On-line
learning during navigation
21
4. Experiments Aspects considered ACIS
Learning convergence Learning convergence
rate Generalization (reactions to environmental
changes) Fault tolerance ACIS performance in
the trajectory optimization problem A
comparison with Hollands classifier systems
(LCS) Hollands classifier system was selected
due to its similarities with ACIS
22
1) Learning convergence of ACIS mouse in a
labyrinth To test if the mouse learned to move
following a food trace Food F Venom V
Scenario 1
Scenario 2
23
(No Transcript)
24
(No Transcript)
25
2) Generalization Test environments
26
(No Transcript)
27
(No Transcript)
28
3) Trajectory optimization
29

• 5. Conclusions and Future Work
• ACIS learning convergence
• ACIS generalization (reactions to environmental
  changes)
• ACIS was able to recognize the modified objects
  in the test environments
• ACIS good performance solving the trajectory
  optimization problem
• ACIS showed a faster convergence than the
  classifier system (LCS), due to the additional
  ACIS affinity maturation process

30

• Future Work
• Implement a global navigation model, where the
  agent creates an internal representation of the
  environment to allow a better exploration of the
  environment
• Add new processing modules to the multilayer
  decision system in order to test the robustness
  of the multilayer approach when new layers are
  added
• Comparison of ACIS with other techniques
• Validation of this research in a real-world
  environment through a hardware implementation

31
8. References 1 Y.H. García. Procesos
markovianos de decisión en el modelamiento de
agentes racionales. Universidad Nacional de
Colombia, Bogotá, Colombia, 2002. 2 Douglas J.
Futuyma. Evolutionary Biology. Sinuaer Asociates,
1986. 3 Carlos Chimal. La experiencia
consciente y la integración de la realidad.
Avance y perspectiva, 20 259- 269, Julio-
Agosto. 2001. 4 R.A. Brooks. Elephants dont
play chess. Robotics and Autonomous Systems,
1990. 5 R.A. Brooks. A robust layered control
system for mobile robots. IEEE journal of
robotics and automation., 1986. 6 E. Charniak
and D. McDermott. Introduction to Artificial
Intelligence. Adison- Wesley, Reading,
Massachusetts, 1985. 7 S. Haykin. Neural
Networks, a comprehensive foundation. Prentice
Hall, second edition, 1999. 8 J. Holland, K.
Holyoak, R. Nisbett, P. Thagard, (1989)
Induction, Processes of Inference, Learning and
Discovery,Massachusetts, USA, MIT Press . 9
Leandro N. de Castro y Jonathan Timmis,
Artificial Immune Systems A New
Computational Intelligence Approach", Ed.
Springer, 2002. 10 F. M. Burnet. Clonal
Selection and After, In Theoretical Immunology,
pp. 63-85, 1978. 11 L. N. de Castro, F. J. Von
Zuben. An Evolutionary Immune Network for Data
Clustering, Proc. of the IEEE Brazilian Symposium
on Artificial Neural Networks, pp. 84-89, 2000.