autopilot 2001
DESCRIPTION
AutoPilot 2001. Jerrold F. Stach, Ph.D. Eun Kyo Park, Ph.D. School of Interdisciplinary Computing and Engineering University of Missouri – Kansas City. Perception by Fuzzy Membership Function. Multi-attribute Decision Making for Agent Mobility. AutoPilot Framework. - PowerPoint PPT PresentationTRANSCRIPT
AutoPilot 2001
Jerrold F. Stach, Ph.D.
Eun Kyo Park, Ph.D.
School of Interdisciplinary Computing and Engineering
University of Missouri – Kansas City
Perception by Fuzzy Membership Function
Multi-attribute Decision Making for Agent Mobility
September 20, 2001
AutoPilot Framework
Sensing
Perception
Reasoning
Behavior
Meta data
•Years 1,2 concentrated on the theoretical basis of mobility and construction of baseline simulator. •Network load leveling was demonstrated as a second order effect of individual agent mobility
decisions.•Year 3 concentrated on quantification of perception using characteristic functions and subjective time.
Autonomous, Rational Agent
September 20, 2001
Sensory Functions
Sensing
Objective Time
Distance
Population Density
Trader Place and Local Service Place Inquiries
•Current Time•Distance in Hops•Queue Length•Arrival Rate•Service Rate
Service Planner at SPprovides instantaneouslocal measures.Trader Place provides measures of remote SPs since last update.
September 20, 2001
Perception In Subjective Time
Perception
•Congestion •Acceptance with Goodness of Fit•Acceptance with Certainty•Difference•Reliability/Mortality
September 20, 2001
Reasoning
Reasoning
•Next Migration•Next Computation•Death (Subjective Time)
- indicates future work
- indicates intermediate progress
September 20, 2001
Behavior
Behavior
Non-Deterministic Choice– stay or go– next location– next computation– self replicate– genetic mutate(signature splice)
•Request Transport
- indicates future work
- indicates intermediate progress
September 20, 2001
Meta Data
Meta data
- indicates future work
•Life History (experiences)•Algebraic Signature (Genotype)•Phenotype•Intermediate Data e.g. progress toward goal, beliefs etc.•Join locations
- indicates intermediate progress
September 20, 2001
2000 Results
Single attribute functions were given for – Distance
(Objective time based on hops and payload)– Cost of Service– Accuracy (quality) of Service
Mobility was solved using a graph theoretic solution which is optimal but has exponential running time
Service Places were weighted in a task graph using a multi-attribute normalization
September 20, 2001
Mapping of Subjective Time to Scalar Time for Linear Attributes such as Cost and Accuracy was Given:
1. Compute the Origin and Limit of Scalar Time Bounds of current network diameter
2. For each attribute:I. compute the slope of the attribute scale
II. obtain the time correspondent
III. compute the mass of the attribute using its weight*Time Correspondent value
3. Create a Time Vector of the attributes
September 20, 2001
Linear and Scalar attributes cont.
4. Compute the mass of the time vector as a multi-body system:
attribute. the
orign to theconnecting vector theis
.AutoPilot)in (as 1 M tosum weights
attribute when the
and generalin 1
th
,
,
r
rmx
rmM
x
qi
qi
September 20, 2001
2000 Observations
Many environmental (sensed) attributes do not scale linearly– congestion– quality– reliability– acceptance
AutoPilots must be able to reason over attributes with various CDFs in subjective time
September 20, 2001
2001 Observation
Many non-linear, environmental attributes exhibit characteristic CDFs over a universe of discourse
– congestion (exponential)
– strength of yes/no (parabola)
– magnitude of difference (logarithmic)
– reliability/mortality (bath tub)
September 20, 2001
2001 Research Goals
Develop a set of relevant perception functions producing Percepts by Fuzzy Membership Functions | 0≤i≤ for Service Place and Service attributes Develop a method to interpret the Percepts for individual attributes
Prove the multi-mass function developed in 2000 is pareto-optimal
Prototype and validate the Percepts
September 20, 2001
The notion of membership
For a “fuzzy” set A→[0,1], A is calledthe membership function and A(u) for u U is called the degree of membership of u in the fuzzy set.
The degree of membership is not intended to convey a likelihood or probability that u has some particular attribute.
September 20, 2001
2001 Research Tasks
Design ways to get “reasonable” membership functions
Functions should have good correspondence to the subjective notions they represent
Functions should be based in theory, i.e. a characteristic function over the universe values of the attribute.
September 20, 2001
The Notion of Perception
An Agent’s life is finite in the system
An Agent carries a Phenotype and Genotype (task signature) yielding an expectation of the duration of work
An Agent must therefore “sense” its own mortality with regard to achieving its goal, i.e. reason in subjective time.
September 20, 2001
Example - Perceiving Congestion
Perception
Safe Region
Unsafe Region
Waiting Time as a Function of Service Place Utilization
The vertical line can be moved to the left according to the agent’s subjective model of time. “Congested”nodes need not be considered in the mobility decision.
September 20, 2001
Example - Perceiving Congestion
Perception
Safe Region
Unsafe Region
Waiting Time as a Function of Service Place Utilization
The vertical line can be moved to the left according to the agent’s subjective model of time. “Congested”nodes need not be considered in the mobility decision.
September 20, 2001
Example - Perceiving Congestion
Perception
Safe Region
Unsafe Region
Waiting Time as a Function of Service Place Utilization
The vertical line can be moved to the left according to the agent’s subjective model of time. “Congested”nodes need not be considered in the mobility decision.
Theoretical Basis
Characterizing the World
September 20, 2001
Trader Place is a Sensor with Memory
At each update interval the following is reported from each Service Place to its Trader Place– Service Place Name < name >
– Node Queue Length Lq
– Agent Service Rate μ– Agent Arrival Rate λ
A Service Place can inquire to the Trader Place <?World> and receive response < {[SP1, Lq,μ,λ],s1,s2,...,sk}, ..., {[SPn, Lq,μ,λ ],s1,s2,...,sm} >
September 20, 2001
Observation
Trader Place update intervals are relatively long compared to agent arrival rates and service rates
Each Trader Place Update is a snapshot of one state of the Universe at a near past instant of measurement
Trader Advertisements are “recent history”, not current state.
September 20, 2001
Agent Sensory Functions
An Agent can enquire to the Service Place <?D,Service_Place_Name> with response <Service_Place_Name,h> where d is in hops.
An Agent can enquire to the situated Service Place <?Environment> with response <Lq,μ,λ> for current local information
An Agent can Inquire to the Service Place <?service_name> and receive reply < [SP1, Lq,μ,λ] ... [SPn, Lq,μ,λ] > where SPn is a Service_Place_Name.
September 20, 2001
Argument for Exponential Streams In The Agent Population
At any observation SP staten can only transition to staten+1 (birth) or staten-1 (death), independent of arrival rate or time. This is the memoryless property of an exponential stream.
Exponential distribution is the limiting distribution of the normalized statistic of random samples drawn from continuous populations
Exponential distribution provides the least information where information content has entropy. It is the most random law and is a conservative approach to modeling the agent population as a dynamic entity as we move to an A-Life model of the AutoPilot agency.
September 20, 2001
Service Place Population Characterization
let be arrivals per unit of time and be services per unit of time.
tn
ne
nt
tp !)(
n
i
ti
n iet
tp0 !
)()( :CDF
September 20, 2001
Service Place State Characterization
Let pn be the percentage of time in steady state the system is in state n.
Assuming the probabilities sum to 1 over the states then
. 1
00if convergeswhich and
0
0 1
np
p
n
p
0
1 1
ondistributiy probabilit the iswheren
pp
n
n
pn
pn
p
pn
September 20, 2001
Service Place Effectiveness
q
q
L
L
L
'
;1
/1/
:Queuesempty -non of Size Expected
:Length Queue Place ServiceMean
:Place ServiceAt Agents OfNumber Expected
/ limit in the Place Service of use Average
1/ )1(0
;1 forn
nppp
September 20, 2001
Service Place Effectiveness continued
0for t )1(
1)(
)(
on.distributiy probabilit cumulative its and queuein time thebe Let
sizemax_queue_ :service refused beingAgent ofy Probabilit
.1)0(:Empty Place Service findingAgent ofy Probabilit
tet
qW
qW
(t)q
Wq
T
p
)/(11:i statecurrent in timeholding Expectediii
/q
Theoretical Basis
The Notion of Fuzzy Sets and Membership
September 20, 2001
The notion of a fuzzy set
A “crisp” set is defined
A x if 0
A xif 1)(x
A
A “fuzzy” subset of a set U is a function
1,0U
On the Powerset P(U) of all subsets of U are the familiar functions of union, intersection and complement.