Presentation at STeP02, Oulu, Finland

1

• Life-Like
• CONTROL

2
Plausibility of cognitive models

• It is easy to define architectures/algorithms for
  special purposes
• Real test for plausibility
  How easily the framework can be
  extended to other applications?
• For example, can an approach developed for
  perception (input) be used for control (output)?

3
Cognitive approach

• Cognitively relevant data consist of clusters
  and linear substructures, data vectors being
  collected in matrix C
• Interpretations
• Rows of matrix C are patterns and features
• Rows of matrix C are categories and attributes
• Rows of matrix C are concepts and their nuances
• In general, rows of matrix C are (numeric) chunks

4
Interpretation among systems

• Interpretation
• Rows of matrix C are operating points and degrees
  of freedom
• Locally linear (affine) state-space system
  models can be represented in the cognitive
  framework

5
(No Transcript)
6
Challenges

• Input-output structure of representations
• Dynamic nature
• Nonlinearities
• Adaptivity
• High dimensionality
• Noise ...

7
1. Input-output structure

• To extend the input oriented model, a strategy
  is needed to implement regression
• Match the known quantities against the internal
  low-dimensional model
• Reconstruct the unknown quantities using the
  internal representation.

8
Associative regression
9
(No Transcript)
10
2. Dynamic nature

• The system state can be compressed into a state
  vector, history can be forgotten given the
  future inputs, the state dictates the future
  behavior

11
3. Nonlinearities

• Clusterwise models facilitate constructing
  piecewise linear models

12
4. Adaptivity

• Model is learned from the observation data
  (slow adaptation)
• Load disturbances, etc., can be learned within
  the model (fast on-line adaptation)

13
5. High dimensionality noise

• Correlation structures are learned spurious
  noise is ignored
• Being based on principal components,
  high-dimensional data is compressed

14
Outlook of chunk control
15
1. Learning behaviors

• Various times, apply some known controller to
  bring the state to the goal
• Learn the combinations of u y pairs into the
  model (GHA algorithm etc.)
• Apply the learned control by associatively
  recalling u when y is measured.

16
Associative control
17
(No Transcript)
18
2. Optimizing behaviors

• Control theory For a linear (affine) system
  quadratically optimal control can be implemented
  as linear (affine) state feedback
• Principle of optimality (dynamic programming)
  Construction of optimal controls can be carried
  out in a modular fashion

19

• Vary control slightly
• If the result is better than before, adapt
  towards the new control
• Random search? - Yes, but still efficient
  The whole (local) model is simultaneously
  adapted, not only the pointwise value

20
Application example
21
(No Transcript)
22
(No Transcript)
23
(No Transcript)
24
(No Transcript)
25
(No Transcript)
26
AI approaches to control