imitation programming larry bull university of the west of england

21
Imitation Programming Imitation Programming Larry Bull University of the West of England

Upload: frederick-snow

Post on 13-Jan-2016

212 views

Category:

Documents


0 download

TRANSCRIPT

Page 1: Imitation Programming Larry Bull University of the West of England

Imitation ProgrammingImitation Programming

Larry BullUniversity of the West of England

Page 2: Imitation Programming Larry Bull University of the West of England

Alan Mathison TuringAlan Mathison Turing

Further research into intelligence of machinery will probably be very greatly concerned with ‘searches’ ….

There is the genetical or evolutionary search by which a combination of genes is looked for, the criterion being survival value ….

The remaining form of search is what I should like to call the ‘cultural search’ … the search for new techniques must be regarded as carried out by the human community as a whole.

Intelligent Machinery 1948

Page 3: Imitation Programming Larry Bull University of the West of England

Culture-inspired Search

Artefacts - left between evolutionary generations [Hutchins & Hazelhurst 1990, ALife II procs.]

Cultural Algorithms - belief space to store knowledge to guide evolutionary search [Reynolds1994, EP III procs.]

Ant Colony Optimization – social insect mechanisms for path finding in graphs

[Dorigo et al. 1996, IEEE SMC-B]

Page 4: Imitation Programming Larry Bull University of the West of England

Imitation: MemesDawkins’ meme [1976] is a unit of

cultural ideas, symbols or practices, which can be transmitted from one mind to another through:◦Writing◦Speech◦Gestures◦Rituals ◦Other imitable phenomena.

Meme = Cultural gene.

Page 5: Imitation Programming Larry Bull University of the West of England

Imitation ComputationSupervised Learning - controller

design through copying of human movement

[e.g., review by Schaal et al. 2003, Phil. Trans.]

Reinforcement Learning - value function approximation sharing between multiple agents [e.g., Price & Boutilier 1999, ML 16 procs.]

Particle Swarm Optimization – boids inspired model of social behaviour

[Kennedy & Eberhart 1995, IEEE NN procs.]

Page 6: Imitation Programming Larry Bull University of the West of England

Unorganised Machines: A-TypeIn the same 1948

paper, Turing presented a general knowledge representation scheme.

Parallel, recurrent networks of 2-input NAND gates.

Page 7: Imitation Programming Larry Bull University of the West of England

Discrete Dynamical Systems

Unorganised machines have a finite number of possible states and they are deterministic.

Hence such networks eventually fall into a basin of attraction.

Page 8: Imitation Programming Larry Bull University of the West of England

Motivation

Discrete dynamical systems are known to (potentially) exhibit an inherent robustness to faults.

Unorganised machines are made from uniform components.

How to design dynamical circuits from such systems through culture-inspired mechanisms?

Page 9: Imitation Programming Larry Bull University of the West of England

Related WorkEvolving Cellular Automata

◦[Packard, 1988][Mitchell et al., 1991][Sipper, 1997]

Evolving Random Boolean Networks◦[Kauffman, 1993][Lemke et al., 2001]

Evolving Graph Representations◦[Fogel et al., 1966][Teller & Veloso,

1996][Poli, 1997][Miller, 1999][Teuscher, 2002] …

Page 10: Imitation Programming Larry Bull University of the West of England

Imitation ProgrammingCreate initial random populationEvaluateFor each individual

◦Select individual in population to imitate◦Copy a randomly chosen aspect (with

error)◦Evaluate◦Adopt new solution if better (or if smaller

but =)

Repeat

Page 11: Imitation Programming Larry Bull University of the West of England

ImitationPick an individual

◦Best used here◦Smallest in case of ties (or random)

Pick a trait at random◦Node start state◦Node connection ◦Network size

Test for error (pe = 0.5 here) which moves connection +/- 1 node id

Page 12: Imitation Programming Larry Bull University of the West of England

SizeImitator is smaller

◦Copy next node in imitatee and connect into network

Imitator is larger◦Cut end node and reassign

connections to itSame size

◦Randomly add a copied node or cut last node

Size limited [inputs+outputs, max.]

Page 13: Imitation Programming Larry Bull University of the West of England

Unorganised Machines as Circuits

1

4

5

6

Input 1

Input 2

Input 3

2

3

Output

Each node encoded as two integers and a binary start state.Nodes are initialised, inputs applied, network run for T cycles,and the output node(s) value is read.

Page 14: Imitation Programming Larry Bull University of the West of England

Multiplexer =20, T=15, N=I+O+30

Page 15: Imitation Programming Larry Bull University of the West of England

a0

a1

d1

d2

d3

d4

out

1

1

0

0

0

0

1

1

0

1

1

0

1

0

1

0

0

0

0

1

1

0

1

1

1

1

0

0

0

1

1

1

Page 16: Imitation Programming Larry Bull University of the West of England

Demultiplexer

Page 17: Imitation Programming Larry Bull University of the West of England

Comparison with EvolutionEP [Fogel et al., 1966] was designed

for graph representations using mutation:

◦ Change output◦ Change transition◦ Add state◦ Remove state◦ Change initial state

One operator used per offspring here.(+’) selection

AB

A C

1/

0/

1/

0/

1/0/

Page 18: Imitation Programming Larry Bull University of the West of England

Multiplexer – IP vs EP

Page 19: Imitation Programming Larry Bull University of the West of England

Demultiplexer – IP vs EP

Page 20: Imitation Programming Larry Bull University of the West of England

Summary of Brief ComparisonImitation always faster and smaller.But selection process in particular

is very different.IP described akin to a population of

hill-climbers.If alter imitation to copy a randomly

created individual stat. same on demux but still better on mux.

Page 21: Imitation Programming Larry Bull University of the West of England

ConclusionsCulture appears to be a somewhat

under-explored source of ideas for AI.

Imitation a ubiquitous process in nature.

Can be viewed as both a component-wise xover operator and mutation.

Initial results suggest competitive with genetic evolution schemes.

Much more exploration needed.