Knowledge Genesis Group

1
Knowledge Genesis Group Smart Solutions
Petr Skobelev
Multi-Agent Technology Ideas, Experiments
and Industrial Applications
Small, but coordinated forces, produce magic.
Prof. A. Konovalov. Lectures on supramolecular
chemistry
Ekaterinburg, 12-13 May 2011
2
Agenda

• Introduction
• Key Challenges of Real Time Economy
• Multi-Agent Technology
• First Experiments with Multi-Agent Solutions
• Industrial Applications in Real Time Scheduling
• Future

3
Knowledge Genesis Group

• Started 1997, Samara, Russia
• Originally from Russian Academy of Science and
  Aerospace Industry
• 15 years of experience in Multi-agent systems
  and Semantic web
• Expertise in application development,
  large-scale systems, web-applications, GPS
  navigation and e-maps, data bases, mobile
  solutions
• 100 J2EE and .net programmers
• Knowledge Genesis Group companies
• Magenta Technology (UK) - 2000
• Knowledge Genesis Germany 2008
• Knowledge Genesis UK 2009
• Emergent Intelligence, USA 2010
• Smart Solutions, Russia 2010
• Advanced technology product vision for solving
  complex problems
• Own development platform
• International network of partners
• Strong links with universities

4
In Samara Office of Magenta Technology (UK)
15 June 1990 The beginning
Prof. George Rzevski (Open University, UK) and
Prof Vladimir Vittikh (Institute of Complex
Systems of Russian Academy of Science)
Company Growth (Number of Employees)
5
Key Challenges of Real Time Economy

• Uncertainty, Complexity Dynamics of business
  are growing
• Clients, partners resources demand more
  individual approach
• High efficiency of business requires to become
  more open, flexible and fast in decision making
• Solutions for Real Time Decision Making can help
  to optimize resources, balance and reduce cost
  time, service level, risks and penalties
• Activity-Based Cost (ABC) model is required to
  analyze options and provide dynamic pricing in
  real time
• Pro-actively negotiate with clients and
  resources on the fly
• Solutions need to support not only optimization
  of resources but also provide opportunities for
  business growth, learning and adaptation
• Use full power of Internet services, GPS
  navigation, mobile phones, RFID, etc

New generation of software solutions for smart
decision making support and sophisticated user
interaction is required on the market!
6
Multi-Agent Technology Differentiation
Traditional Systems
Multi-Agent Systems

• Networks of agents
• Parallel Processing
• Negotiations Trade-Offs
• Distributed
• Knowledge-Driven
• Self-Organization
• Evolution
• Thrive with Complexity
• Managing growth

• Hierarchy of programs
• Sequential Processing
• Top-down instructions
• Centralized
• Data-driven
• Predictable
• Stable
• Reduce Complexity
• Full Control

Modules are working as a co-routines
simultaneously
7
Distributed Approach Wins!
8
The Beginning of Multi-Agent Systems

• Started in the beginning of 1970s
• Based on achievements in Artificial Intelligence
  Object-Oriented and Parallel Programming
  Telecommunications
• Traditionally focused on logic reasoning
  (Wooldridge, etc)
• Our approach is bio-inspired (Van Brussel, Paulo
  Letao, etc) but strongly influenced by
• Ilya Prigozhin in Physics (auto-catalytic
  reactions),
• Marvin Minsky in Psychology (society of mind),
• Artur Kestler in Biology (holonic systems)
• Key focus self-organisation and evolution,
  synergy, non-linear thermodynamics, collective
  (emergent) intelligence
• First Applications Internet e-commerce
• Current Applications logistics, data mining,
  text understanding, etc
• Future Web-Intelligence

9
Classification of Agents
 
Current Focus
10
How Agents work?

• A new agent is created at runtime whenever there
  is a task to be performed
• The agent begins its life by analysing the task
  and studying rules of engagement
• Agent activities include
• analysing situation
• composing messages
• receiving sending messages to other agents or
  humans
• interpreting received messages
• deciding how to react
• acting upon their decisions
• This enables agents to run concurrently
• When an agent completes its task it is destroyed

11
Examples of Multi-Agent Systems

• Winestein Technologies http//www.weinstein.com
• NuTech http//www.nutech.com
• Living Systems http//www.livingsystems.com
• AgentBuilder - http// www.agentbuilder.com
• Quarterdeck - http// arachnid.qdeck.com
• GeneralMagic - http//www.genmagic.com
• Intelligent Reasoning System
• - http//members.home.net80/marcush/IRS
• BiosGroup http www.eurobios.com
• LostWax http//www.lostwax.com
• About 30 companies on the market.
• More than 100 University projects are known.

12
Existing Multi-Agent Systems

• Single-Agent Approach no self-organization
• Based on results of traditional AI research
  (Prolog-style deductive machine for reasoning)
  not effective for dynamical environments with
  high uncertainty
• Concept of Mobile Agents problems with security
• Traditionally Oriented on e-Commerce
• Do not have Knowledge Base and Reasoning Tools to
  support Decision Making Processes of End-Users
• Do not have Re-Negotiations support
• Memory intensive and slow low performance, only
  a few Agents can work on server in parallel
• Not supported with development tools (basic
  platforms only) - very expensive and difficult to
  design develop

13
Our Approach Main Ideas

• Our Multi-Agent Systems working in Swarms
  consisting of a large number of small autonomous
  programs (objects) called Smart Agents
• Smart Agents have special in-built tools for
  decision making and ontology-based scene support
• Main feature of Smart Agents is the ability to
  solve complex problems through negotiations
• Every complex problem can be solved by
  self-organization and evolution, in competition
  and cooperation of Smart Agents
• Examples real time logistics, pattern
  recognition, text understanding, data mining, etc

14
Demand and Supply Matching on Virtual Market
Engine - is Core Part of Real Time Multi-Agent
Solutions for Any Type of
Complex Problems
Swarm Demand and Supply Networks
Virtual Market Engine
S
S
Demand-Supply Match
D
D
S
S
MatchContract
D
S
S
D
D
D
S
Demand Agent
SupplyAgent
S
S
D
S
D
S
D
D
15
Our Approach Main Ideas

• Software Agents model physical objects, people
  and abstract concepts forming Virtual Markets in
  which they allocate supply to demands
• The agent interaction is based on the free-market
  model Demand Agents purchase resources from
  Supply Agents and Supply Agents sell resources to
  Demand Agents, all working concurrently
• Logistics orders to resources
• Text understanding words to meanings
• Data mining records to clusters
• Agents learn how to accomplish their tasks by
  accessing Ontology where they consult the
  detailed knowledge of the domain in which they
  work

16
Ontologies

• Ontology by definition is knowledge as it is or
  conceptualisation of abstraction (Gruber)
• Knowledge can be represented by semantic network
  of concepts and relations
• OntologyEditors OntoEdit, WebODE, WebOnto,
  Protégé-2000, OWL/RDF/RDQL
• Our ontologies are used for pecification of
  situations (scenes)
• Ontologies are the combination of declarative
  semantic network and operational knowledge
  (scripts)
• Concepts and relations can represent objects,
  roles, properties, processes, attributes, etc.
• Some ontologies can be hard-coded to improve
  performance

17
Example of Ontology /Scene
18
Multi-Agent Platform

• Adaptive, Real time and Event-driven
• Swarm-based approach (vs mobile agents)
• Virtual Market as a Core engine
• Highly Reactive Pro-Active
• Provide Emergent Intelligence

• Based on Semantic web innovations
• Ontology to capture Enterprise Knowledge and keep
  it separately from source code
• Decision Making Logic instead of rules
• Able to Learn in Future (Using Pattern Discovery
  module) - source codes separated from knowledge

MAS driven by Inner virtual market
Knowledge BasedDecision Making

• Java-based / .net
• Peer-to-peer architecture
• Scalable/Robust
• Strong visualizations
• Desk-Top Web-Interface

EnterprisePlatform
19
Smart Clash Analysis for Airbus Wings

1. When change happen (lets assume that in our
   example it is size of part C) we create agent of
   changed part
2. This agent will investigate scene and find his
   neighbors (Part B)
3. Agent of part C will create agent of part B and
   to inform him on changes and his new boundaries
4. Agent of Part B will compare new size or position
   of Part C and will check his boundaries according
   with the type of relation
5. If these changes not affect his position it
   will be recognized as the end of wave of changes.
   
6. If yes the situation will be repeated for other
   neighbors of the network in the same way (Part A,
   Part E, Part D)
7. As a result of this process it will be
   ripple-effect from initial change which will take
   place until it affects positions of other parts

Semantic network scene of wing
Part E
Is-Assembly
Part A
Part D
Fix-Link
Fix-Link
Agent of Part C
Part B
Can-Rotate
1, 2
Part C

• Value for Client
• Analysis can be made in real time (and even for
  dynamically reconfiguring complex objects)
• The approach proposed can be applied for any
  complex object or machine without full
  re-programming (need another ontology and scene
  mainly and change of interpretation of links)
• Many threads of activities can go in
  quasi-parallel mode starting from any point
  (changed part) of network when needed and even in
  parallel for engineers

3
4
Agent of Part B
Ripple effect of Part C changes
20
Noble Group Solution Smart Coal Mining in
Indonesia

1. When change happen (lets assume that in our
   example it is bad Weather in Region A heavy
   rain!) we create agent of region and send message
   about weather event.
2. This agent will use ontology for find out the
   consequences. Usually bad weather affects Jetty
   Loading Rate and Waterway availability. Then
   agent finds all affected instances of jetties and
   waterways in this region and inform them about
   bad weather.
3. Agents of all these objects will estimate impact
   and make changes in their schedules. This can
   leads to new changes in a network. For example
   due to the new jetty schedule some barge will be
   late on anchorage. Agent will inform his
   operator immediately and will present current
   options, for example barge will come 3 days late.
4. If this decision will be confirmed by operation,
   Agent of Barge will create agent of Vessel and FC
   and will inform them about delay, and they will
   check options and inform their operators if
   needed.
5. If these changes will not be possible to solve
   inside region and they will affect client it
   will be needed to inform clients and it will be
   the end of wave of changes.
6. As a result of this process it will be
   ripple-effect from initial change which will take
   place until it affects positions of other parts

Agent of Weather in Region A
Semantic network scene of Noble Group network
1
Region A
Is-Contain
2
Client 4
Jetty 1
Booked
Contract
Barge 2
3
4
Booked
Booked
Crane 4
Vessel 3

• Value for Client
• Team of managers can be coordinated in real time
  according with events coming
• The approach proposed can be applied for any team
  coordination without full re-programming (need
  another ontology and scene mainly and change of
  interpretation of links)
• Flexibility many threads of activities can go in
  quasi-parallel mode starting from any point
  (changed part) of network when needed and even in
  parallel for users

Ripple effect of changes
21
Jarvis Solution Smart Pattern Recognition
Agent of House

1. Input image flow comes as binary digital photos
   taken on new landscapes with different
   configuration of patterns and high level of
   noise.
2. All agents of patterns start their work in
   parallel and compete because it is not known in
   advance where strong patterns will be recognized.
3. Looking into ontology all agents trying to make
   their best match with image fragments (and all of
   them can invoke some specific methods for this).
4. If for one of patterns matching is Ok then he
   adds object into scene specifying parameters of
   recognized pattern (lake, forest, etc) and links
   it with other objects.
5. If matching is not Ok (for example agents of
   house and cloud have conflict and are competing
   for the same fragment of image in brackets)
   they need help and switch for cooperation based
   on domain semantics.
6. For this example agent of house will look in
   ontology and find out that usually there are
   garage and road near houses. Now he can
   investigate scene and will see that garage and
   road are already there.
7. Then probability of the fact that it is house
   (not a cloud) becomes higher because of this
   links (sometimes it is needed that garage and
   road will agree also that their neighbor looks
   like a house).
8. In this situation cloud also can know from
   ontology that it can move and will give priority
   to the house.

Agent of Cloud
Agent of Lake
Agent of Garage
Agent of Forest
Agent of Road

• Value for Client
• Analysis can be made in real time or batch image
  processing
• The approach proposed can be applied for any
  complex image processing system for pattern
  recognition without full re-programming (need
  another ontology and scene mainly and low level
  methods of image processing)
• Flexibility of solution many threads of
  activities can go in quasi-parallel mode starting
  from best recognized parts of image (unknown in
  advance)
• Quality of pattern recognition can be very high
  because of semantic links and errors checking
  during the process of recognition
• Proposed solution is very generic not only for
  image processing but also for text understanding
  and other applications (patterns of sense can
  compete for strings of texts, etc)

Semantic network partially reconstructed scene
during patterns recognition
House
Placed-Near
Between
Garage
Lake
Road
22
OmPrompt Solution Smart Fax Recognition

1. This task has the same solution as for images
   considered above.
2. When new fax is coming agent of first pattern
   according with fax template starts looking his
   part of image. If he finds 100 matching he
   writes results in scene and initiates next agent
   looking into scene of fax template.
3. But if matching is not 100 a few agents of this
   area can compete for the same part of the image
   (for example Osipemagen it is wrong end of one
   field and wrong beginning of the new field).
4. Agent of first field will recognized that it is
   beginning of the address and will ask agent of
   the next field do you recognize rest of the
   string as a company name connected with this
   address? In general the best one will try to get
   support from other with whom he can cooperate
   investigating his local area via relations.
5. Recognized part of image is saved in scene and is
   used by all other agents to detect next parts and
   find solution of conflicts.

Real fax
Is-Header on the top of page
Fax Number
Next in line
From
To
Below
Below
Name
Semantic network partially reconstructed scene
of fax recognition

• Value for Client
• Analysis can be made in real time
• The approach proposed can be applied for any
  complex fax or image processing without full
  re-programming (need another ontology and scene
  mainly and change of interpretation of links)
• Flexibility many threads of activities can go in
  quasi-parallel mode starting from any point
  (changed part) of network when needed and even in
  parallel
• Quality of fax (image) recognition can be very
  high because of semantic errors checking during
  the process of recognition

Fax template
23
VineWorld Solution Smart Diet Management

1. When new event happen (lets assume that in our
   example it is user request to replace Fish by
   Pork at dinner time) we create agent of changed
   object
2. This agent of Pork will replace Fish informing
   other agents in dinner group and agent of dinner.
   
3. Immediately agent of white wine (good with fish)
   will leave the dinner and agent of red wine will
   propose Dinner agent to enter the menu as a good
   match with user preferences.
4. Agent of Dinner will calculate calories and find
   out that now it is more than 2000 calories for a
   day.
5. To solve the conflict agent of Dinner will try to
   find candidates to reduce number of calories
   calculating the difference.
6. If it is not possible to solve the conflict
   inside dinner it will ask agent of Tuesday menu
   who else can be involved in this process. Maybe
   both other groups (lunch and breakfast) will be
   recommended to start looking variants in
   parallel.
7. All potential candidates will be asked to find
   nearest possible food option according with user
   preference and less calories.
8. All options will be not simply sorted and
   presented to user for final decision but will
   compete to be recognized as a best option. Best
   possible option (remove ice-cream) can also
   switch to cooperation with other options to get
   more points.
9. As a result of this process a few food items can
   drop out of menu, or size of portion will be
   reduced or physical exercises will be added to
   menu to reduce extra calories.
10. In all cases it will be ripple-effect from
    initial change which will take place until
    decision is found or not

Semantic network Scene of Tuesday menu
Apple juice 177 kcal
Omelet 261 kcal
x
Breakfast
ltemptygt
Soup 205 kcal
Lunch
Pudding 362 kcal
Ice cream 450 kcal
! 50
Strawberry 41 kcal
Pork 537 kcal
River fish 216 kcal
Dinner
Red Wine 180 kcal
White Wine 192 kcal
Total 2225 kcal
Total 1904 kcal
Total 1988 kcal
Ice cream 50
Agent of Menu
Refuse omelet
Change Wine
!

• Value for Client
• Solution can be find in real time (and even
  during update of food items types)
• Solution is open for adding new types of
  services health, exercises, fridge, etc
• Solution is flexible changes can start from any
  point and run in parallel threads of activities

Bicycle
24
Smart Content Semantic Network of Celebrities
25
Upload and specify new photos
26
Ontology of Celebrities
27
Ontology/Scene Editor
28
Add new photo and agents will change network
29
New Photo is added to Semantic Network
30
Text Understanding Projects

• Intelligent Documents Classifier (Rubus/Aon)
• Classification of all documents into groups with
  the similar sense - semantic proximity
• Ability to build the template document on the
  base of the group of similar documents
• Intelligent Requests System (Integrated Genomics)
• Intelligent search and comparison of the
  abstracts semantic descriptors on the basis of
  the problem domain ontology
• Database Natural Language Requests System (Hotel
  Booking)
• Intelligent partial matching on the base of the
  ontology to make complex search of several
  interconnected items
• On-line clustering analysis of customers types
  and their patterns of requests thus generating
  new rules to enlarge the ontology

31
MEDLINE Database - Internet search for
molecular biology abstracts
The MedLine database contains brief abstracts of
articles on biological themes, which are
presented to users free of any charge.
Search conditions - keywords and logical
expressions
If the abstract of a found article is satisfies
the user, he can order the full version of the
article for a certain price.
32
Text Understanding Process
Syntax stage
Example phrase MagentA will provide support for
Software Programs employed by the Client.
Morphology stage
Semantics stage
33
Text Understanding System
Two pUC-derived vectors containing the
promoterless xylE gene (encoding catechol
2,3-dioxygenase) of Pseudomonas putida mt-2 were
constructed. The t(o) transcriptional terminator
of phage lambda was placed downstream from the
stop codon of xylE. The new vectors, pXT1 and
pXT2, contain xylE and the t(o) terminator
within a cloning cassette which can be excised
with several endonucleases.
34
Text Understanding System
Two pUC-derived vectors containing the
promoterless xylE gene (encoding catechol
2,3-dioxygenase) of Pseudomonas putida mt-2 were
constructed. The t(o) transcriptional terminator
of phage lambda was placed downstream from the
stop codon of xylE. The new vectors, pXT1 and
pXT2, contain xylE and the t(o) terminator
within a cloning cassette which can be excised
with several endonucleases.
35
Text Understanding System
Two pUC-derived vectors containing the
promoterless xylE gene (encoding catechol
2,3-dioxygenase) of Pseudomonas putida mt-2 were
constructed. The t(o) transcriptional terminator
of phage lambda was placed downstream from the
stop codon of xylE. The new vectors, pXT1 and
pXT2, contain xylE and the t(o) terminator
within a cloning cassette which can be excised
with several endonucleases.
36
Text Understanding Systems
Results In good groups in general accuracy of
finding correct article is higher than 81, in
certain requests its almost 90 In bad groups
the probability of still good article put there
by mistake is less than 8
Intelligent Requests System statistics Time to
build one semantic descriptor 1-2 min. Time to
search through 1000 abstracts 1 min. Ontology
of problem domain contains 150 concepts and
3100 relations (with inheritance)
37
Text Understanding System
Comparison with keywords The proposed approach
demonstrated significant quality increase
comparing to keywords Keyword search even with
all improvements (synonyms etc) still
demonstrates rather bad results, clearly
insignificant to the required task Accuracy of
proposed search higher than simple keyword search
Intelligent Requests System statistics Time to
build one semantic descriptor 1-2 min. Time to
search through 1000 abstracts 1 min. Ontology
of problem domain contains 150 basic concepts
and relations
38
Multi-Agent Solutions for Real Time Resource
Allocation, Scheduling and Optimization
Your solution application?
39
MAT Solutions for Real Time Logistics

• Truck Scheduling
• Ocean Scheduling
• Taxi Scheduling
• Courier Scheduling
• Car Rental Optimization
• Factory Scheduling
• Airport Scheduling
• Work forces ...

40
How It Works in Transportation Networks
VOL 10 PALLETS SLA 5 DAYS
VOL 10 PALLETS SLA 10 DAYS
VOL 5 PALLETS SLA 2 DAYS
20
20
60
VOL 10 PALLETS SLA 10 DAYS
VOL 5 PALLETS SLA 8 DAYS
20
60
80
120
20
60
It is important to be able to assess alternate
routes, to meet services levels and minimum cost.
60
100
40
Imagine the power of having a single system that
can automatically plan and re-plan a network like
this, as events occur, such as new orders being
added or resource availability changes.
This order has a shortest journey route but the
capacity is not available on one of the legs.
41
Transport Logistics Network Complexity

• Real-time scheduling with shrinking time windows
• Large complex networks (gt 1000 orders per day,
  gt 100 locations, gt 50 vessels )
• Less-than-Truck loads requiring effective
  consolidation
• Need to find backhaul opportunities
• Intensive use of crossdocking operations
• Trailer swaps
• Numerous constraints on products, locations, dock
  doors, vehicles types, availability,
  compatibility
• Individual Service Level agreements with major
  clients
• Own and third-party fleet
• Fixed and flexible schedules
• Dependent schedules (trailers, drivers, dock
  doors, etc)
• Activity Based Cost Model
• Other client-specific requirements

Most of large complex transport networks are
still scheduled manually!
42
Vision of MAT Scheduling Solutions
Current Situation and Ongoing Plan
Advise on How-To make Network More Efficient
Pattern Discovery
Evolutional Design
Patterns and Ongoing Forecast
Resulting Plan and KPIs
Adaptive Scheduler
Input Events Flow (New order, Resource
unavailable, etc)
Network Configuration Situation specs (Scene)
Domain Ontology
Ontology Editor
Simulator
Network Designer
Network Assets Real Situation
Modeling Plan and KPIs
Domain Knowledge
Modeling Data (Flow of orders, fleet size, etc)
43
MAT Schedulers Screens Example
44
Ontology as a Way to Capture Domain
Knowledge
Describe your classes of concepts and relations
45
Examples of Concepts and Relations
Ontology concepts

• Client
• Order
• Cargo
• TI
• TIConsolidation
• Fleet
• Trailer
• DD Trailer
• Standard
• Truck
• Tractor
• Rigid

• Dock
• Trip
• Location
• Cross Dock
• RDC
• TI Operations
• Collect
• Drop
• Truck operation
• Stop
• Move
• Idle

Ontology Relations

• ClientHasOrder
• OrderHasCargo
• OrderHasTI
• FleetHasTruck
• FleetHasTrailer

• TruckHasSchedule
• TIConsolidationHasTI
• JourneyHasTI
• TIHasTISchedule
• TIHasTIOperation

46
Truck Logistics Scene Example

• Scene objects
• 27 clients
• 154 cargoes
• Fleet
• 22 DD Trailer
• 12 Rigid Truck
• 72 locations
• MANCH
• MILTO
• EXEBOTFR
• CHIPP
• CONIC
• YORFI
• PENRITFR

Create a situation (scene)
47
Logic of Multi-Agent Scheduling

• Existing schedule
• New Order arrives
• Pre-matching
• New order wakes up Truck 3 agent and starts
  talking to him
• Truck 3 evaluates the options to take New order
• Truck 3 wakes up Order 3 agent and asks it to
  shift
• Order 3 analyzes the proposal and rejects it
• Truck 3 asks New order if it can shift to the
  right
• Truck 3 decides to drop Order 3 and take New
  order
• Agents of New Order and Truck 3 disappear
• Order 3 starts looking for a new allocation and
  finally allocates on Truck 1 by shifting Order 1

0800
1600
12.00
2000
No
Time
Truck 1
My time window is too tight I cannot shift
Truck 2
Can you shift to the right?
Truck 3
Back
Next
48
Logic of Multi-Agent Routing
Cross dock 1
Cross dock 2
Y
Consider business-network of a company 1.Order1
goes from Location C to Location Z 2.Order2 goes
from Location B to Location X 3.Order3 appears,
which goes from Location A to Location Z 4.Order3
decides to go to B and then travel with Order 2
via cross-dock1 5.Order4 appears, which goes from
Location A to Location Y 6.Order3 decides to
travel the first leg with Order 4 and the second
leg with Order 1 via cross-dock 2, to avoid going
alone from A to B
Back
Next
49
Case Study UK Logistics Operator

• Network Characteristics
• 4500 orders per day
• Order profile with high complexity
• Many consolidations should be found
• Few Full Truck Load orders
• Few orders can be given away to TPC
• Majority of orders require complex planning the
  price of a mistake is high
• 600 locations
• Large number of small orders
• 3 cross docks
• 9 trailer swap locations
• 140 own fleet trucks, various types
• 20 third party carriers
• Carrier availability time
• Different pricing schemes

• Problems to be Solved
• Location availability windows
• Backhaul
• Consolidation
• Vehicle capacity
• Constraint stressing
• Planning in continuous mode
• Dynamic routing
• Cross-docking
• Handling driver shifts

Key Problem Real-time planning in a highly
complex network with X-Docks and Dynamical
Routing
50
Case Study UK Market leader in supply chain
management

• Network Characteristics
• Employs around 5,000 staff, rising to 7,000
  during Christmas peak
• Has 40 operating sites
• Manages 300,000 sq m of warehouse space
• Has sites across Europe
• Has a turnover of 400 million
• Moves in excess of 10bn worth of merchandise
  each year
• Services over 3,000 retail outlets around the
  globe
• Travels 75m miles each year
• Operates a fleet in excess of 1,300 vehicles
• Has over 35 years of supply chain experience

• Problems to be solved
• Maximise utilisation of capacity minimise need
  for ad hoc journeys
• Comply with constraints temperature regimes,
  collection and delivery times, customer priority,
  product compatibilities, product weight, etc
• Optimise trunking through best use of changeovers
  and cross docks
• Do not over split orders to prevent problems on
  reconsolidation
• Make best use of subcontractors versus own fleet
• Make best use of store returning vehicles

Key Problems Automatic Search for effective
scheduling decisions using own fleet Adaptively
distribute orders among the journeys of static
schedule
51
Summary of Benefits (Before / After)
BEFORE IMPLEMENTATION
AFTER IMPLEMENTATION
Two operators worked for a day to make a schedule
for 200 instructions
8 minutes to schedule 200 transportation instructi
ons
Planning day 1 for day 3 no chance to Support
backhauls and consolidations in real time
Planning day 1 for day 2 and even day 1 for day 1
No software for schedule 4000 orders With
X-Docks and Drivers (manual procedure only)
4 hours to plan orders 4000 orders via X-Docks
and ability to add new orders incrementally (a
few seconds for a order)
Knowledge was hard to share, it was spread
among different experts
Capture best practice and domain knowledge in
ontology. New knowledge can be inserted quickly.
Hard to consider various criteria quickly and
choose the best possible option
Choosing the best route from the point of view
of consolidation or other criteria
52
Case Study Taxi Dispatching (UK)

• Network Characteristics
• Call centre with about 130 operators receiving
  orders concurrently
• A fleet of more than 2,000 vehicles (each with a
  GPS navigation system)
• A very large number of orders more than 13,000
  orders per day the order flow occasionally
  exceeds the rate of 1,500 orders per hour order
  arrival times and locations are unpredictable
• The order attributes are as follows place of
  pick-up and drop urgent or booked in advance
  (for a certain date and time) type of service
  (minivan, VIP, etc.) importance (a number from 0
  to 100 depending the client) special
  requirements (pet, need for child chair, etc.)
• A large variety of clients, e.g., personal,
  corporate, VIPs, with a variety of discounted
  tariffs, with special requirements for drivers,
  disabled, requiring child seats, requiring
  transportation of pets, etc.
• A large number of freelance drivers who lease
  cars from the company and are allowed to start
  and finish their shifts at times that suite them,
  which may differ from one day to another
• At any time around 700 drivers are working
  concurrently, competing with each other for
  clients
• Guaranteed pick up of clients in the centre of
  London within 15 minutes from the time of placing
  an order
• Unpredictability of the traffic congestion in
  various parts of London causing delays and
  consequently the interruption of schedules,
  unpredictability of times spent in queues at
  airports and railway stations

Key Problem Real-timeresource reallocation

• Problems to be solved
• React on events in real time
• Provide individual approach to clients
• Balance costs vs time and risks
• Increase efficiency of business
• Satisfy drivers

53
Case Study UK Corporate Taxi Company

• Main Results
• The system began its operation and maintenance
  phase in March 2008, only 6 months from the
  beginning of the project
• The total number of processed orders increased
  7 (1000 orders per day 20 pounds cash in
  average) in a first month with the same number of
  resources
• 98.5 of all orders were allocated automatically
  without dispatchers assistance
• The number of lost orders was reduced to 3.5 (by
  up to 2 )
• The number of vehicles idle runs was reduced by
  22.5
• Each vehicle was able to complete two additional
  orders per week spending the same time and
  consuming the same amount of fuel, which
  increased the yield of each vehicle by 5 7
• Profitability Increase 4.8
• Orders collecting time 40 faster
• Time for Operators Training 4 times less
• ROI 6 months

54
Key Customers

• Avis (UK) Leading car rental provider
• Real time scheduler for downtown market reducing
  car assets required and improving service levels
• Addison Lee (UK) largest private hire car firm
  in London
• Operational system and real time scheduler for
  resource optimization
• Tankers International (UK) Manage a large oil
  tanker fleet
• Real time scheduler for tankers scheduling and
  optimization
• One Network (USA) logistics software provider
• Providing development services to implement new
  core, scheduling and visual features/components
  for their platform
• GIST (UK) MS supply chain
• Real-time scheduler for increased fleet
  utilisation and reduced transportation costs
• Airbus/Cologne University (Germany)
• Catering RFID scheduler for improving service
  level and airport efficiency
• Enfora (USA) major manufacturer of handheld
  devices
• Development of a wide range of software modules
  and market partnership for a real time scheduling
  web service
• Aerospace Enterprises Energy,CSKB-Progress,
  Izevsky Motozavod (Russia)
• Prototyping P2P network of real time workshop
  schedulers for workers optimization
• RusGlobal Prologics (Russia)
• Real time truck scheduler for resource
  optimization
• Russian Fund of Fundamental Research, Ministry of
  Science and Education

55
Adaptive Network of Real Time Schedulers
16.00 Monday
10.00 Monday
Event 2 delay in truck delivery
Event 1 delay on factory side
Real Time Factory MAT Scheduler
Real Time Truck MAT Scheduler
Enterprise Service Bus
XML messages
Event 1 Factory is late for 4 hours with
producing products. Factory Scheduler need to
negotiate with Truck Scheduler on re-scheduling
previously booked truck and avoid penalties for 4
hours delay. If booked truck can be reallocated
by Truck Scheduler to other client no penalties
is required.
Event 2 Now small Truck is late. Then Truck
Scheduler need to negotiate with Factory
Scheduler that bigger truck (more expensive) will
be sent to Factory to avoid penalties. Factory
Scheduler can re-schedule production lines to
produce more products which can be loaded into
this truck for the same client to use capacity of
bigger truck fully.
56
Future
That Was Then
This is Future
Real-time
Batch
Manage Trade-offs
Optimizers
Decision-Making Logic
Rules Engines
Cost/value equation
Constraints
Learn, Simulate Adapt and Forecast
Visualize
57

• Enjoy beauty of self-organized systems for
  solving complex problems.
• Thank you!

58
Knowledge Genesis Group Smart Solutions
Petr Skobelev
Multi-Agent Technology Ideas, Experiments
and Industrial Applications
Small, but coordinated forces, produce magic.
Prof. A. Konovalov. Lectures on supramolecular
chemistry
Ekaterinburg, 12-13 May 2011