Engineering Conscious Machines Ricardo Sanz Autonomous Systems Laboratory Universidad Politécnica de Madrid Models of Consciousness ESF/PESC Exploratory

Engineering Conscious Machines

Ricardo Sanz

Autonomous Systems LaboratoryUniversidad Politécnica de Madrid

Models of ConsciousnessESF/PESC Exploratory Workshop

BirminghamSeptember 1-3, 2003

2R. Sanz / Engineering Conscious Machines

Abstract There are several ongoing attempts to build conscious

machines using different types of technologies and engineering methods: conventional symbolic AI, neural networks, non-linear dynamical systems, etc.

If a computer-based mind for a machine is going to be made as conscious as humans, it is quite clear that it will be a complex software/hardware application.

There are several approaches to engineering complex software systems and they will be reviewed in this talk with a particular emphasis on constructive methods (obviously due to a bias of the speaker). Emergent methods will be also analysed and convergent technology introduced.

We will analyse the possibilities of the different methods an try to extrapolate from present software engineering technology the degree of effort needed, the time scope and the tradeoffs of different designs and implementation technologies.

A Clear Need

(of machine consciousness)


Simple feedback

SensorsSensors

Physical MachinePhysical Machine

ActuatorsActuators


Enhanced feedback

SensorSensor


ActuatorActuator

FilterFilter ExecutorExecutor

Goals


Stateless vs Stateful


State

SensorSensor ActuatorActuator


Goals


Layering


State



StateFilterFilter ExecutorExecutor

Layer 1

Layer 2

Goals


Deliverative/Reactive


State



StateFilterFilterReasoning

EngineReasoning

EngineKnowledge


Model-based control


State



ModellerModeller ExecutorExecutorWorldModel


Introspection and Reflection


ActuatorActuator

StateFilterFilterReasoning

EngineReasoning

EngineKnowledge

SensorSensor

Meta-level representationQueryEngineQueryEngine


ControllerControllerControllerController

Hierarchy and heterarchy





ControllerController

ControllerController

ControllerController ControllerController ControllerController


Multiresolutional reflective control Revonsuo: “multiple levels of organisation,

forming a hierarchical, causal mechanical network”

Industrial controller evolution is mimicking biological mind evolution

Now, we’re in the phase of creating conscious controllers (even when most control engineers don’t know)


Autonomic Systems(as IBM sees them) Adapts to changes in its environment Strives to improve its performance Heals when it is damaged Defends itself against attackers Exchanges resources with unfamiliar systems Communicates through open standards Anticipates users’ actions

Possesses a sense of selfSciAm May 06, 2002


Positioning Need of Need of Artificial ConsciousnessArtificial Consciousness

A A foreignerforeigner point of view point of view A A customercustomer point of view point of view A A reductionistreductionist point of view point of view

In summary: an In summary: an engineerengineer point of view point of view


Structure of the talk Give a personal perspective on Give a personal perspective on present present

state of affairsstate of affairs

Give an personal assessment on Give an personal assessment on engineering processesengineering processes for for ACAC

Summarize a personal, half-baked, Summarize a personal, half-baked, engineering vision engineering vision onon UTC UTC

Give a personal perspective on Give a personal perspective on present present state of affairsstate of affairs

Give an personal assessment on Give an personal assessment on engineering processesengineering processes for for ACAC

Summarize a personal, half-baked, Summarize a personal, half-baked, engineering vision engineering vision onon UTC UTC

State of Affairs

A Personal Perspective


Some Views Taylor: “Attention is the gateway to

consciousness” Salichs: “Basic consciousness: Current

(here and now) self” Holland & Goodman: “Consciousness via

incremental intelligence” Sloman & Chrisley: “Architectures that

support mental processes conected with normal notion of consciousness”

Cleermans: “flexible, adaptive control over behavior”


CODAM “The corollary discharge of attention module

(CODAM) carries the signal of ownership of the about-to-be-experienced contentful activity arriving later on the sensory input buffer from sensory feedback.”

J.G. Taylor, From Matter To Mind

Translation into conventional control tech: CODAM fills Measure’s “Owner” slot with “I”

Sensor: X31 Value: 2.44 Owner: I Time: 13:23

Sensor Measure


CODAM Splitting of attention control signal:

to focus sensorial attention and to create the PRS waiting for sensor inputs

CODAM is a simple sensor control architecture It seems complex due to all that confusing

decorations based on neurophysiological details about functional topology of human brains

CODAM generates a “self”, i.e. a continuously refreshed dynamical model of the agent


CogAff Consciousness is not just a bag of information

processing functions. It is a core architectural principle of minds (with a demonstrated evolutionary advantage).

There is a difference between an architectural schema like CogAff (a pattern) and an agent architecture like H-CogAff (an instance)

There exist reusable mind design patterns


First Assessment We share a single core engineering model

of consciousness (while unconsciously)

Neuroscientific/psychological data corroborates this model

Varying visions are just views of this core model coloured of personal backgrounds


The modeling machine Evolution has engineered a model-based Evolution has engineered a model-based

learning controller: learning controller:

the the Central Nervous SystemCentral Nervous System

This machine generatesThis machine generatesdifferent typesdifferent types of of models to act models to act in the worldin the world


Naïve models of reality Judging that an animal will not mind being killed if

it is not offended, Eskimos take various ritual precautions before, during, and after the hunt.

The rationale (the behavioral model of the world+agent) lies in the belief that animal spirits exist independent of bodies and are reborn: an offended animal will later lead his companions away so that the hunter may starve.

Just projections of what is best known: the self


Survivor Models Elementary, ad-hoc, Elementary, ad-hoc, experience-experience-

based causal modelsbased causal models of reality of reality Examples: agriculture, mating, Examples: agriculture, mating,

micronesian navigationmicronesian navigation (rowing to move the islands (rowing to move the islands to certain positions in theto certain positions in thehorizon)horizon)

CleermansCleermans: : “representational “representational systems that can be systems that can be adaptively modified adaptively modified by ongoing experienceby ongoing experience”


Deep models Behavior based on deep models outperforms

behaviour based on behaviourally learnt models

Scientific theories of reality


The machine of the world Science and technology have established

themselves as the best models and tools to control the machinery of the world

Engineering Processes

How can we build these systems ?


Processes Alternative processes:

Theory-based construction (Sloman) Incremental hacking (Holland) Reverse engineering brains (Redgrave) Learning from raw stuff (Aleksander) Emergent/autopoietic growth (Doran/Ziemke)

What’s contingent and what’s necessary? What’s the core design pattern of

consciousness ?


Beyond “Normal” Agents Dependable control agents do have

requirements well beyond what is considered “normal” intelligent function:

Real-time performance Embeddability High-assurance Evolvability / Ugradeability


Theory

Design-based processes

Design

Artefact

Chrisley: ”Flow of effect is not one-way” Holland: ”toy around and keep eyes open”


Complex Systems Engineering Doran: “Our ability to design and build complex

agent architectures is limited”

Round-trip engineering is not possible due to sybsystems interaction explosion

Learning / adaptation methods are useless due to design-space complexity explosion

Architecture-based product families and modularity can help a little

Design for emergence is a promising alternative


The Way Learn from evolution:

From model-based behavior to model-based engineering

Theory-based, model-implemented, tool-supported engineering processes for mind construction

Plenty of examples out there

Personal Vision

What else ?


WorldWorld

ManMan

Man on His World


Meaning generation Autonomous systems:

Generate meanings from data (typically from sensory inputs)

Use their continuosly updated mental models to control behavior

Meanings are equivalence classes of agent+world trajectories in state-space in relation with agent’s value system (projections into the future including counter-factuals)

Hesslow: “simulated perception can be elicited by (simulated) behaviour”


Example 1 Consider:

a nuclear reactor the primary control system the primary protection system

What’s the meaning of a particluar measure of neutronic flow ?

The’re two different meanings for the control system for the protection system

futurenow

supercritical

subcritical

critical

protection thresold


Example 2 Consider that you’re Owen driving along

Bristol Road Consider the meaning of a road sign

saying “Manor House to the right”

If Owen becomes aware of the sign, the value of his future along Bristol Road changes


Awareness and Consciousness A system is aware if it is generating meanings

from perceptions (including inner perceptions)

A system is conscious if “I am aware” is valid in the the present state of affairs (it is generated from the perceptual flow, i.e. the system is aware of itself).

Lacombe: “It is impossible to separate awareness, consciousness and understanding”.


Strong points of this model Unifying Machine applicable Explains other related phenomena: e.g.

attention Of what do you calculate potential effects

when resources are scarce? Only of those pieces that most assuredly can

affect your future: focus of attention Holland: “simulate only the part of the world

that can mostly affect the agent”


Another Strong Point This model provides a metric

It is possible to calculate the degree of coverage of future trajectories

It makes possible the comparison of awareness levels of systems in the same conditions (i.e. experiencing the same sensor space, including inner space)


Synthesising (Control) State A model-based, sensor processing + control

hierarchy reduces data (by abstraction+integration) until reaching a single, unique, compact representation of the self

This self-generation process is mostly hidden and hence the self seems directly perceivable and inmaterial


Synthesising (Control) State Manzotti: “uniqueness is important for

consciousness” Haikonen: “inmaterial mind … missing

perception of material carrying symbols or processes”

Anceau: “conscious behaviour is sequential”


Summary Mind is a multiresolutional phenomenon of control

Minds generate and use dynamic models

At any resolution level, meaning generation generates awareness

At any resolution level, mind reflection generates consciosuness

We -usually- only can talk about the upper level

Thanks for being conscious

during my talk!

Questions ?

Documents

Engineering Conscious Machines Ricardo Sanz Autonomous Systems Laboratory Universidad Politécnica de Madrid Models of Consciousness ESF/PESC Exploratory