Immunity in self-aware systems

Professor Jon TimmisDepartment of Computer Science

Department of ElectronicsYork Centre for Complex Systems Analysis

ImmunityPlants and animals have immune systems to help protect and repair

Images: Mark Coles

Swarming, FlockingApparent co-ordination of many individuals - no central control

bbc.co.uk

Collective BehaviourCollectives work together to solve problems that one (or a few) can not

manage on their own

bbc.co.uk

Co-ordination

Fire flies appear to synchronise across a swarm

youtube.com

“Swarm” immunity?immune cells acting together in an individual?

Survival as a collective?

bbc.co.uk

An engineering perspectiveBuilding systems that operate for long periods of time without human

intervention is a challenge

Big dog

A swarm of robots

Building systems that can cooperate to perform tasks an individual can not do is also a challenge

boston dynamics; swarmanoid

So reality is different to sci-fi!

An immune system for a robot?

Sense and react to events external to the robot

Sense and react to events internal to the robot and swarm

Sense and react to internal and external events in an organism

An immune system sensing and reacting to the environment

We live in hazardous environments. T cells are one line of defence.

There are many hazardous environment we might like to monitor.

Sniffer dog robot with an immune system for sensing and reaction

The sniffing robot

An  immune  system  for  a  “swarm”  of  robots?

Simple  tasks  ...

13

Foraging• Define behaviour as a FSM

shaped environment with object placed randomly within the

arena. Each robot is programmed with behaviour-based sub-

sumption architecture [1]. With this architecture, behaviour

modules for a foraging task include obstacles avoidance,

moving to object, grab object, move to base, deposit, scan

arena and random walk. These behaviours can be expressed

with a finite state machine and is given in Fig. 4. A simulation

starts with robots dispatched from the base to explore within

the arena to find, collect, and deposit attractors continuously

until end of simulation or until all energy used. Initial energy is

set to 15000 unit and energy consumption for each action per

second is as follow: avoidance 0.9, randomWalk 0.8, moveTo-

Food 0.8, grabObject 1.2, scanArena 0.8, depositObject 1.2

and moveToBase 0.8 unit.

grabObject moveToObject

depositObject

avoidance scanArenamoveToBase

leaveBase randomWalk

ObjectInGripper

Gra

bS

uccess

AtB

ase

ClearClear

ObstacleDetected ObstacleDetected

Done InArena

ObstacleDetected

Obje

ctI

nS

igh

t Ob

jectL

ost

Obje

ctI

nS

ight

ObstacleDetected ObstacleDetected

Fig. 4. Finite state machine for behaviours of a robot in a foraging task.

B. Experiments Settings

Fixed Parameters for Simulation. Table I is the settings

for fixed parameters in this study. The values for these parame-

ters are set purely by arbitrary choice and can be changed if de-

sired for other studies. Estimated time for a robot to collect and

deposit one object τ̂ � (arenawidth+arenaheight)/2motorspeed � 10

0.15 s.

To collect all 3000 object by 15 normal robots in optimal

case, an estimatedtotalobjecttotalrobot x τ̂ �13333.33 is required but

for completeness this number is rounded up to 13500 giving

the simulation duration. For our current work, we only study

anomaly detection for single robot failure but future work will

consider failures on multiple robots.

TABLE I

FIXED PARAMETERS FOR ALL SIMULATIONS.

Parameter Value Parameter value

Arena size 12 m x 12 m Robot normal speed 0.15 m/s

No. robot 10 unit Component fault Motor (Wheels)

Static OPR 0.15 unit/s Simulation duration 13500s

Max. object 400 unit Fault injection time 5000s

Initial object 100 unit Sampling period 250s

Total object 3000 unit No. faulty robot 1 unit

Variable Parameters for Simulation. We are interested in

detecting anomalies caused by component faults in the robot

under various environmental conditions. These conditions are

time-varying and thus the behaviour of the SRS changes

accordingly, i.e., normality changes. We are interested in

detecting anomalies under these conditions and analyse the

effects it has on the detection accuracy. Table II shows the

time-varying environmental conditions, the types and magni-

tudes of the component faults.

TABLE II

VARIABLE PARAMETERS FOR VARIOUS SIMULATION SCENARIOS.

Parameter Value

Environment static, varying object distribution VOD ,

varying object placement rate VOPR, vary-

ing presence of obstacles VPO

Fault type complete Pcp, partial Ppt, gradual Pgr

Fault magnitude partial 0.045m/s to 0.125m/s,

gradual -0.00005m/s to -0.001m/s

In this paper, we analyse the effect of these environmental

conditions against the detection accuracy of RDA utilising

relative normality (rnRDA). In addition, we also look at the

sensitivity of the detection against different magnitudes of

failures.

• Env.VOPR:. Effect of time-varying object placement rate

(availability of object) on the detection accuracy.

• Env.VOD:. Effect of time-varying object distribution (lo-

cation of object) on the detection accuracy.

• Env.VPO:. Effect of time-varying presence of obstacles

on the detection accuracy.

• F.Ppt:. Detection accuracy with different magnitudes of

partial failure.

• F.Pgr:. Detection accuracy with different magnitudes of

gradual failure.

Parameters of rnRDA. In this study, the input data to the

rnRDA is the vector of 3 variables <OBJ, ENG, DIST>

where OBJ={obj1,obj1, ...,objm} is the number of object

collected, ENG={eng1,eng1, ...,engm} is energy usage, whilst

DIST={dist1,dist1, ...,distm} is distance travelled within 250s

and m is the number of neighbor at particular time. For

each variable, 20 equally spaced receptors scaled from 0 to

maximum expected value are used. Total stimulation for each

receptor in Eq. 1 is calculated using Gaussian kernel with

kernel width h[obj]=1, h[eng]=8, and h[dist]=2. For other

parameters, β=0.01, b=0.1, gb=1.1, and α=1. As in Eq.5,

anomaly is detected when receptor position rtp(x) ≥ l for

any variable.

IV. DISCUSSION ON RESULTS

Performance evaluation. The performance of the detection

algorithm is evaluated based on true positive rate TP , false

positive rate FP , true negative rate TN and false negative rate

FN . Accuracy of detection is represented by

Accuracy =(TP + TN)− (FP + FN)

TP + TN

Simula;on  of  foraging

State  diagram  for  robot  behaviour

A bit more complicatedKeeping a swarm of robots together

State diagram to control a robot

Things can go wrongMaybe not as robust as we first thought?

Collective repairThe immune system has many responsesWhich we can develop computer models ofWhich we can exploit in swarm robots

An individual immune system that monitors everyone else as well

Swarm  to  Organism

SYMBRION

Organisms

Self-Assembling “Swarm” failure Self-Assembling

“Organism” failureRecovery after assembly

What about underwater?

Creating healthy robots and swarms

Made some initial advances, but we are a long way from really reliable robots and swarms that can operate autonomously

Make use of insights from immunology to help drive our technology

Can a robot have an immune system?Yes, but not the same as you and I !!

Acknowledgements

• University of York, Royal Society, European Commission, EPSRC, BBSRC, MRC, Dstl

• Kieran Alden, Paul Andrews, Iain Bate, Lynette Beattie, Mark Coles, Richard Greaves, James Hilder, Pete Hickey, Rita Ismail, Paul Kaye, Vipin Kumar, Kelvin Lau, Tiong Lim, Alan Millard, Lachlan Murray, Daniel Moyo, Mark Neal, Jenny Owen, Nick Owens, Fiona Polack, Mark Read, Susan Stepney, Andy Tyrrell, Alan Winfield, Richard Williams, Eva Quarnstrom

A bit of fun