kitteninterestAI and Robotics

Nov 15, 2013 (2 years and 11 months ago)


Producing a



Group Discussion Outbrief

Guy Lemieux, Daniel Coore

Dagstuhl 06361


Global to local

Local to global

Middle ground?

Ixodes Ricinus

Global to Local

These global objectives and behaviours
can be “compiled down” into locally
specified solutions

Coordinate systems


List homomorphisms

Boundary value problems

Questions about G to L

How should these be manifested in a programming
language ?

How should we go about finding more of them?


Complex Systems Community

Should the global behaviour specification be vague?

Are there multiple correct global behaviours (non

Does the global algorithm have to be completely specified?
(imperative versus declarative)

Is it agnostic of actual embedding of the processors?

Local to Global Issues

Can we infer/prove global outcome, or do we have to simulate?

Can we restrict L to make inference possible?

If restriction allows us to characterize trajectory in “robust” way, answer
is yes/maybe (could be difficult)

It may be chaotic with attractor states (good and/or bad)

Example: PDE solving is a form of “local specification/behaviour”
(difference equations) that are run to produce a global outcome

Is Local to Global a convex optimization problem?

If yes, then outcome may be easily predictable

Is this a robust control problem?

Maybe this is how we restrict L to make inference possible

Can this property be used to compile a global specification into local

… merging G to L

with L to G …

aka “The Conclusion”

“G to L” vs. “L to G” …

How about “G to M to L” instead?

Add an interface “M” between G and L

This may give both G and L something to concretely
refer to as a middle ground

Like an instruction set architecture merges an
algorithm/language to the microarchitecture

Do we need an M ?

The G
L relationship is a recursive hierarchy

G can be replaced by G
L to give G

L can be replaced by G
L …

Possible Primitives for M

Primitive (vote count to question)

Question: Is this “easy” to do in your “spatial computing paradigm”?

Gradients (7)

Majority votes (7)

Movement transaction (3)

Atomic messages (maybe too low level) (6)

Compartmentalization (across ensemble) (5)

Reduction (maybe too high level) (1)

Maintaining connectedness (1)

Scatter/gather (0)

Globally unique ID (0)

Almost globally unique ID (4)

Can be achieved by generating a large random number at each node