A Rationale for Artificial Consciousness

Autonomous Systems LaboratoryUniversidad Politécnica de Madrid

Agrigento, Italy, 23-24 November 2005

Ricardo Sanz

International Workshop on Artificial Consciousness

The very first

Thanks to the organisers for inviting me to this workshop

Apologise for using this opportunity to present bold, half baked ideas and ongoing work behind our research

Contents

Introduction to complex control

Approaches to autonomy

Modelling approach to development

Model-based self-aware control systems

Introduction

Content

This speech is a reflection, that describes some core ideas behind long term research in complex control systems

The objective of this research is to provide architecture-centric, component based technology for the construction of high autonomy control systems

from dreams ...

Forbidden Planet

(1956)

Robbie

2001: A Space Odyssey

(1968)

HAL

Star Wars

(1977)

C3POR2D2

Blade Runner

(1982)

RoyPris

ZhoraLeon

Terminator

(1984)

T800

I Robot

(2004)

Sonny

... to realities

Field

Process

Business

Continuous Process Plant

Fieldbus

Control Network

Enterprise Network

ProcessControl

MIS

Data Storage

Process Operation

Field Configuration

Sensing and Acting

Safety

Chemical Plants

Computerised cars

Adaptor

Body Electronics Network

Adaptor

I/O

Adaptor

Suspen-sion

I/O

Adaptor

SteeringManager

I/O

Adaptor

PowerTrain

Adaptor

GatewayBody

AssistantSystem

DriverInterface

I/O

I/O

Adaptor

BrakeManager

Distribution Grids

www.robot.md

... like Dr. Susan Calvin in “I, Robot”

The Problem

Science and technology of computer-based control is facing an enormous challenge when increased levels of autonomy and resilience are required from the machines that are supporting our ambient

We can say, without doubt, that control systems complexity is boosting

Are Systems Too Complex?

In some sense, software intensive controllers are becoming too complex to be built by traditional software engineering methods.

Two main effects:

Increase in size

Decrease of dependability

Complexity forces change

E01-EDI

Data Warehouse(Interfaces to and from the

Data Warehouse are notdisplayed on this diagram)

G02 - GeneralLedger

A05 - AP

S01 - SalesCorrections

I01 POReceiving

I03 Return toVendor

I06 WarehouseManagement

S06 - Credit App

P15 EES EmployeeChange Notice

OTHER APPS - PCAP - Collections/Credit

TM - Credit Card DB

ACCTS REC APPS - PC990CORBad Debt

Beneficial FeesBeneficial Reconcile

JEAXFJEBFAJEBKAJEDVAJESOAJEVSAJEVSFNSF

TeleCredit Fees

INVENTORY CONTROL APPS - PCCode Alarm

Debit ReceivingsDevo Sales

Display InventoryIn HomeJunkouts

Merchandise WithdrawalPromo Credits

RTV AccrualShrink

AP Research - Inv CntrlAP Research-Addl Rpts

Book to Perpetual InventoryClose Out Reporting

Computer Intelligence DataCount Corrections

Cross Ref for VCB DnldsDamage Write Off

Debit ReceivingsDFI Vendor Database

Display Inventory ReconcileDisplay Inventory Reporting

INVENTORY CONTROL APPS - PCDPI/CPI

IC BatchingInventory Adj/Count CorrectInventory Control Reports

Inventory LevelsInventory Roll

Merchandise WithdrawalOpen ReceivingsPI Count Results

PI Time Results from InvPrice Protection

Sales Flash ReportingShrink Reporting

SKU Gross MarginSKU Shrink Level Detail

USMVCB Downloads

Journal Entry Tool Kit

Scorecard - HR

L02-ResourceScheduling

P09 - P17Cyborg

M02 - Millennium

M03 - Millennium 3.0

Banks - ACH and Pos toPay

Cobra

B01 - StockStatus

S03-Polling

P14 On-line NewHire Entry

CTS

Plan Administrators(401K, PCS, Life,

Unicare, SolomonSmith Barney)

D01 Post LoadBilling

I04 HomeDeliveries

I02 -Transfers

Arthur Planning

I07 PurchaseOrder

I12 EntertainmentSoftware

I05Inventory Info

E13E3 Interface

S04 - Sales Posting

V01-Price ManagementSystem

I10 Cycle PhysicalInventory

I55 SKUInformation

K02Customer Repair

Tracking I35 Early WarningSystem

B02 MerchandiseAnalysis

I13- AutoReplenishment

U18 - CTO

Intercept

I09 Cycle Counts

E02-EmployeePurchase

Texlon 3.5

ACH

Stock Options

I17 Customer PerceivedIn-Stock

U16-Texlon

SiteSeer

C02 - CapitalProjects

F06 - FixedAssets

US Bank ReconFile

Star Repair

EDICoordinator

Mesa Data

NEW SoundscanNPD Group

AIG Warranty Guard

Resumix

Optika

Store BudgetReporting

P16 - Tally Sheet

Cash Receipts/Credit

S05 - HouseCharges

Ad Expense

L01-PromoAnalysis

V02-PriceMarketingSupport

BMP - Busperformance Mngt

StoreScorecard

I11 PriceTesting

Valley Media

P09Bonus/HR

I15 Hand ScanApps

Roadshow

POS

S08 - VertexSalesTax

A04 - CustRefund Chks

Equifax

ICMS Credit

CellularRollover

S09 - DigitalSatelliteSystem

NPD,SoundScan

Sterling VANMailbox (Value)

I18SKU Rep

X92-X96Host to AS400

Communication

S02 -Layaways

Washington,RGIS,

Ntl Bus Systems

V04-SignSystem

I14 Count CorrectionsNARM

P01-EmployeeMasterfile

I06 - CustomerOrder

FrickCo

UAR - Universal AccountReconciliation

DepositoryBanks

S07 - CellPhones

S11 - ISPTracking

AAS

Fringe PO

Cash Over/Short

L60 MDFCoop SKU Selection

Tool

SKUPerformance

SupplierCompliance

1

I35 - CEIASIS

Misc Accounting/Finance Apps - PC/NTCOBA (Corp office Budget Assistant)

PCBS(Profit Center Budget System)Merchandising Budget

AIMSMerch Mngr ApprovalBatch ForecastingAd Measurement

AIMS Admin

AIMSReportingAd

Launcher

V03- MktReactions

SpecSource

CTO2

RebateTransfer

SignSystem

CopyWriter'sWorkspace

ELTPowerSuite

StoreMonitor

AIS Calendar

Stores & Mrkts

Due Dates

Smart Plus

InsertionsOrders

BudgetAnalysis Tool

Print CostingInvoice App

AIS Reports

BroadcastFilter

Smart PlusLauncher

GeneralMaintenance

Printer PO

PrinterMaintenance

VendorMaintenance

Vendor Setup

Connect 3

Connect 3Reports

Connect 3PDF Transfer

Spec SourceSKU Tracking

S20-SalesPolling

Prodigy

PSP

In-HomeRepair

WarrantyBilling

System

Process Servers(Imaging)

Adaptation as solution

A possible path to the solution of this increasing complexity problem is adaptation

Adaptation can be seen from many perspectives, e.g.

adaptation during implementation and

runtime adaptation.

Examples

A paradigmatic example of the first type (construction) is reusable component retargeting to adapt it to a particular execution platform.

A paradigmatic example of the second type (run-time) is fault-tolerant control.

Can Consciousness help?

Can artificial consciousness be a solution to the performace/dependability problem?

The Obvious

What do we prefer?

A conscious brain surgeon

An unconscious brain surgeon

The reason: conscious real-time adaptation

The Obvious ?

What do we prefer?

A conscious robotic brain surgeon

An unconscious robotic brain surgeon

A-priori rationale

It looks like it is better if machines are conscious

But:

Machines are not humans

Beware !

We must be aware of the differences between sensing and perceiving

Perception is tinted with the colours of phenomenology, that seems to be a private, agent specific, issue (in biosystems)

We would need some form of alloperception for machines

Motivations for AC

Artefacts like us: consciousness, emotion anf affect, experience, imagination, etc. (Robotics)

Studying natural systems with computer laboratory models (Cognitive Science)

Proficient machines (Intelligent Control)

The Focus of our work

Software intensive controllers for autonomous robust behaviour based on self-awareness

Qualia left for future endeavours

Autonomy Approach

Autonomy Approach

An alternative approach is to move the responsibility for correct operation into the system itself.

That means moving the adaptation from the implementation phase into the runtime phase.

During runtime the system perceives changes and adapts to these changes to keep the mission assigned to it during the design phase.

Many Alternatives

Several alternatives are being explored based on the implementation of architectural mechanisms for self-organisation and/or self-repair.

These systems are built and started in a base state and they follow adaptive life-cycles based on the circumstances of the environment that surrounds the computing system (e.g. fault-tolerant systems, automatic learning, genetic programming or autonomic computing)

Autonomic Computing

IBM autonomic computing initiative:

”Autonomic computing is the ability of an IT infrastructure to adapt to change in accordance with business policies and objectives. Quite simply, it is about freeing IT professionals to focus on higher-value tasks by making technology work smarter, with business rules guiding systems to be self-configuring, self-healing, self-optimizing, and self-protecting.”

Autonomic Element

Autonomic Systems

Engineering Models

Model-driven DevelopmentA recent strategy to build complex embedded systems is model-driven development (MDD).

In the words of Jonathan Sprinkle:

[MDD is] “A design methodology used to create and evolve integrated, multiple-aspect models of computer- based systems using concepts, relations, and model composition principles to facilitate systems/software engineering analysis of the models and automatic synthesis of applications from the models.”

MDD ProcessDeep and multiple system models are built and analysed using tools provided by the modelling environment.

These models are later used to (automatically) generate the final system:

This means that the models must necessarily capture the semantic aspects of the final system and not just the structural properties.

MDD Approach

Behaviour Models

Processing Models

Hardware Models

Model Building

Tools

Model Analysis

Tools

System Generator

Configuration Generator

Runtime System

MODELS

GENERATORS

RUNTIME

ENVIRONMENT

DESIGN

ENVIRONMENT

Model-based Systems

Feedback control

ProcessController

Process output

Controlsignal

Commandsignal

Perception-control-action

Perception

Environment

Action

Autonomous Agent

Control

Sample: RCS Model

World ModelPerception

Environment

Action

Autonomous Agent

Value JudgmentFour core elements:

Sensory PerceptionWorld ModellingValue Jugdment

Action Generation

Sample: Model-reference control

Reference Model

Adaptation Mechanism

ProcessController

Reference model output

Process output

Controlsignal

Controlparameters

Commandsignal

ym(t)

y(t)

From design to runtimeA truly adaptive system (a system with the autonomic properties) must necessarily do a semantic processing of the information it has about itself.

This is the type of reasoning that MDD tools help perform over their models if they are semantically precise enough.

Tools/Models that are being developed to be used by builders at the implementation phase can find their “autonomic” use by the system itself at the runtime phase.

Raising SelfEngineering models constitute the very self-model of the running system

The (auto, allo) value system comes with the model

Perception

Environment

Action

Autonomous Agent Body

WorldModel

Model Content

Model Exploitation

World Self Others

Reflective model-based

Classic adaptive control, or fault-tolerant control use plant models to adapt the control to new plant conditions. Fault-tolerant computing uses computation models to keep the computation ongoing.

Reflective, model-based autonomous systems will have deep model-based introspection capabilities that will let them achieve a high degree of adaptability and resilience.

Model ExecutionModels (of the world, of the self, of others) may be executed over virtual machines for different purposes:

Action calculation

Diagnosis and reconfiguration

What-if (imagination)

Retrodiction and causal analysis

Self? For what?

Self-Configuration

Self-Optimisation

Self-Healing

Self-Protecting

Self-Awareness

Self-X

Self-X functionality exploits models of the system itself in the performance of model-based action

The Engineering Vision

Break the barrier between the engineering process and the product

Bridge the gap between construction and run-time

Enable life-cycle adaptation

Make the system self-aware using software self-models + real-time reflection

ConclusionsThe rise of control system complexity claims for new technologies

Adaptation is a key issue. Self-adaptation is needed for many reasons (particularly dependability)

Model technology is critical technology for complex controllers

Model-based control can be extended to the control system itself

Conscious controllers will exploit self models embedded in world models to maximise mission-level effectiveness.

What is it like to be a model-based reflective predictive controller?

“There may be something it is like to be such a self-model linked to such a world-model in a machine with a mission”

A question remains:

Thanks for the Consciousness