Upload
diane-york
View
217
Download
0
Embed Size (px)
Citation preview
19.4.2023 1
Artificial Intelligence in Real-time Systems
LAP 8780 and ISP 9010
Tallinn University of Technology
Professor Leo Motus
19.4.2023 ©L.Motus, 2004 2
J. McCarthy “What is Artificial Intelligence” (November 2004)
science and engineering of making intelligent machines, especially computer programs; need not confine itself to methods that are biologically observable.
Intelligence is the computational part of the ability to achieve goals in the world
AI research started after WWII. Alan Turing’s lecture in 1947 – he was the first to decide that AI was best researched by programming computers rather than building machines
http://www.formal.stanford.edu/jmc/whatisai/
19.4.2023 ©L.Motus, 2004 3
Schools of thought in AI (1)
Conventional AIo Expert systemso Case based reasoningo Bayesian networkso Behaviour based AI
Computational Intelligenceo Neural networkso Fuzzy systemso Evolutionary computation
19.4.2023 ©L.Motus, 2004 4
Schools of thought in AI (2)
Conventional AI – behaviour based AI
A methodology for developing AI based on modular decomposition of intelligence (e.g. Rodney Brooks):
o Robotics and intelligent agents (real-time dynamic systems able to run in complex environments
Computationally leads to interaction-based model of computation, e.g. super-Turing computation
See the course ISP 0012 – software dynamics
19.4.2023 ©L.Motus, 2004 5
Schools of thought in AI (3)
Computational intelligence – evolutionary computation
Applies biologically inspired concepts, e.g. population, mutation, survival of the fittest. These methods divide into two:o Evolutionary algorithms, e.g. genetic algorithms
o for search and optimisationo Swarm intelligence, e.g. ants
o A collective behaviour in decentralised, self-organised systems (e.g. multi-agent systems)
19.4.2023 ©L.Motus, 2004 6
Examples of artificial intelligence based techniques (1)
Basic (algorithm-centred) techniques stem from studying:o Representation of shallow and deep knowledgeo Reasoning (problem solving), including the pattern
(or condition) matching problemso Learning and adaptation (supervised and/or
unsupervised)o Search (including data mining)
By combining the basic techniques more complex problems can be solved – e.g. computer vision
The above-listed techniques are based on imitating processes applied by biological creatures.
19.4.2023 ©L.Motus, 2004 7
Examples of artificial intelligence based techniques (2)
Expansion of the domain where AI techniques were applied, and deeper understanding of the essence of “intelligence” has lead to non-algorithmic techniques:o Agents, info-bots, nanobots, etco Coalition of agents, multi-agent systemso Proactive components, social intelligence (?)
J. Ferber (1999) Multi-agent systems, Addison-WesleyR. Brooks (1986) “A robust layered control system for a
mobile robot”, IEEE J.of Robotics and Automation
19.4.2023 ©L.Motus, 2004 8
Biological paradigms for Artificial Intelligence and Real-time Control
Stem from the functioning principles of humans and other biological species:o Hypothetical division of functions between left and
right hemisphere o A functional model of human brain by Newell and
Simono Studies in swarm intelligence, and animal behaviouro Studies and experiments in molecular biology
19.4.2023 ©L.Motus, 2004 9
Opposing characteristics of the co-resident brain computers
von Neumann serial processor (symbol processing) is believed to operate in the left hemisphere of a human brain
Associative parallel processor (pattern processing) is believed to operate in the right hemisphere of a human brain
19.4.2023 ©L.Motus, 2004 10
Comparison of functions of the hemispheres (human brain)
The computation and/or reasoning is:
in the left hemisphere in the right hemisphere
- linear - non linear
- time sequential - time independent
- batch oriented - multi-tasking
- stacked interrupts - random parallel execution
- word/symbol oriented - pattern oriented
- non-intuitive - highly intuitive
- structured memory - associative memory
19.4.2023 ©L.Motus, 2004 11
Comparison of functions of the hemispheres (human brain)
The computation and/or reasoning is: in the left hemisphere in the right hemisphere
- cumulative correlation - instantaneous multiple correlation
- incremental learning - non-sequential learning- sensory dependent - sensory independent
V. Rauzino, “ Some opposing characteristics of the Co-resident Brain Computers” Datamation, 1982, vol. 28, no.5, 122-136
19.4.2023 ©L.Motus, 2004 12
Newell-Simon functional model of a human brain
Motory actions
Env
iron
men
t
CognitionPerception
Env
iron
men
t
Buffers
Sensors
InterpreterCognitive processor
Internal memory l/s
External memory
Buffers
Human muscles
19.4.2023 ©L.Motus, 2004 13
Basic difference between conventional AI and AI in RT systems (1)
Technical or natural
system System based on AI
1. 2.
3.
4.
Conventional AI is explicitly human centric !
19.4.2023 ©L.Motus, 2004 14
Basic difference between conventional AI and AI in RT systems (2)
Technical or natural
systemSystem based on AI
A1
A2
H1, H2
H3, H4
Humans have just a role of a supervisor in AI applications in Real-time systems !
19.4.2023 ©L.Motus, 2004 15
A view on a real-time system
Environment
A system comprising humans, computers, etc
Task i
Task 1 Task 2
Task n
Task 3
19.4.2023 ©L.Motus, 2004 16
A closer view on a task in a real-time system
Each task can be carried out by applying different methods, e.g.:o Methods based on “natural intelligence “ – i.e.
manuallyo Methods based on Science (e.g. mathematics,
control theory, etc)o Methods based on “artificial intelligence” – i.e. crisp
theory based reasoning, approximate methods of reasoning (e.g. neural nets, fuzzy logic), distributed intelligence methods (e.g. agents)
19.4.2023 ©L.Motus, 2004 17
Intelligent methods (+)
Natural and artificial intelligence based methods are good since they:o Provide efficient solution to a many computationally
complex problemso Decrease the burden of mathematical modellingo Enable to use approximate non-linear methods for
reducing the dimensionality of input spaceo Are capable of drawing unexpected conclusions
and applying unconventional methods on spot.
19.4.2023 ©L.Motus, 2004 18
Intelligent methods (-)
Natural and artificial intelligence based methods are not always applicable because:o Only probabilistic estimates are available for the
quality of obtained solutions (they are approximations of the “scientific” solutions)
o Time for obtaining a solution is indeterminate (the case of deduction based methods)
o Due to insufficient educational background those methods are too often handled as “black boxes” – hence no guaranteed result
19.4.2023 ©L.Motus, 2004 19
Intelligent methods – the case of conventional AI applications
Many independent tasks are solved simultaneously, or rather a single task at a time
The environment cannot influence task execution process – truth values are independent of time and events, occurring in the environment or in the other tasks
Frequent use of backtracking – task execution time is indeterminate
Goals and sub-goals of tasks are static, and are to be fixed before the execution of the task starts
19.4.2023 ©L.Motus, 2004 20
Intelligent methods – the case of AI methods in real-time systems
Many, inter-dependent tasks are to be solved simultaneously (forced concurrency)
The environment can influence the task execution process – truth values may change dynamically, depending on time and events occurring in the environment
Time for execution of a tasks is often strictly limited Goals and sub-goals for tasks may be determined
dynamically (during the task execution) – only a strategic goal is usually fixed before the execution starts
19.4.2023 ©L.Motus, 2004 21
Names used for AI methods are not self-explaining and straightforward
Most of the methods and tools used have historical names and are in-between of pure deductive and pure inductive methods.
For instance, expert systems:o The first-order predicate calculus is a typical expert
system and represents a classical deductive approacho First-generation expert systems (e.g. the frustrated
banker) are a typical inductive approacho Second generation expert system (a mixture of deep
and shallow knowledge) are in between the two approaches
19.4.2023 ©L.Motus, 2004 22
Quality of a task’s solution
In conventional AI application quality means logical and quantitative correctness of a solution – normally a vector comprising, e.g. precision, risk estimate, cost, etc.
In AI application in a real-time system timeliness is added as the highest priority component of the quality vector
Conventional quality-wise – more promising are deductive methods
Time-wise – more promising are inductive methods
19.4.2023 ©L.Motus, 2004 23
Intelligent methods – deductive approach
Paradigm -- top-down approach;
from a general case to a specific caseo humans build a non-contradicting theory, based on
deep knowledge and experimental data o Specific problems are stated (usually by humans) as
special cases of this theory, and then solved by computers
Examples: theorem provers, structural synthesis of programs
19.4.2023 ©L.Motus, 2004 24
Intelligent methods – inductive approach
Paradigm – bottom-up approach;
from a specific case to a general case o Humans provide meta-theoryo Based on meta-theory and a set of examples
(problems with solutions), computers (or humans) build specific theories that resolve a class of problems
Examples: neural nets, inductive synthesis of programs
Note: induction and co-induction
19.4.2023 ©L.Motus, 2004 25
Comparing deductive and inductive approaches
Advantages:Deductive methods provide guaranteed quality of the solution, if obtainedInductive methods have short and rather deterministic execution time
Disadvantages:Deductive methods have indeterminate solution time and high resource requirements (labour-consuming)Inductive methods have usually unknown quality of the solution, formation of the learning set is not easy, learning time is lengthy (building a special theory)
19.4.2023 ©L.Motus, 2004 26
Approximate reasoning (1)
Pragmatic goals:o to obtain interim result in the reasoning process
before any given deadlineo be able to continue reasoning if time and other
resources permit
Implicit assumption – the quality of the reasoning outcome (and interim results) improves proportionally with the given time and resources
19.4.2023 ©L.Motus, 2004 27
Approximate reasoning (2)
Also known as: imprecise computing, any-time algorithms, progressive
reasoning, etc.Basic idea – to make reasoning results available in time-
deterministic way, and to continue reasoning if additional time becomes available
See, for instance, I.R. Chen “On applying Imprecise Computation to Real-
time AI Systems”, The Computer Journal, vol.38, no.6, 1995, ,434 – 442 (kataloog lugemisvara)
Reflex-based approach – a way out for real-time systems?
19.4.2023 ©L.Motus, 2004 28
Approximate reasoning (3)
A simple example of approximate reasoning – forecasting the trends based on observations:o Based on recursive computation of a posteriori
probability densities o Based on recursive adjustment of membership
functions (possibilities), related to many-valued logic and case-base reasoning
Approximate solution methods (Bayesian neural nets and possibilistic neural nets) are used to reduce computational complexities.
19.4.2023 ©L.Motus, 2004 29
Two different clusters of data for computing a posteriori distribution
19.4.2023 ©L.Motus, 2004 30
Approximate a posteriori probability density computed by Bayesian NN
19.4.2023 ©L.Motus, 2004 31
Scattering is used instead of probability density (possibilistic neural net)
19.4.2023 ©L.Motus, 2004 32
Possibility distribution as computed by a possibilistic neural net
19.4.2023 ©L.Motus, 2004 33
Reflex-based approach to approximate reasoning (1)
Imitates the behaviour of biological creatures acting in hard real-time – e.g. car driving, collective games (karate, dancing), riding a bicycle
Observation – an experienced driver makes complex and high quality decisions in a short time, a novice driver cannot reach such decisions (even if given unlimited time)
Hypothesis -- decisions and actions of humans in hard real-time are based on reflexes rather than on conventional reasoning
19.4.2023 ©L.Motus, 2004 34
Reflex-based approach to approximate reasoning (2)
Reflex-based approach to reasoning:o should combine deductive and inductive
approacheso leads not necessarily to an approximation of the
inference treeo creates shortcuts on the inference tree by modifying
inference rules, a set of axioms, or both
A weak similarity – with time deterministic case-base reasoning method as used in the BRIDGE project
19.4.2023 ©L.Motus, 2004 35
AI applications in Real-time systems (examples)
Navigation tasks, computer vision related tasks, performance of AUV, etc
On-line assessment of strategies, generation of alternative strategies and/or goals
Coordinated work of multiple agents, especially time-aware agents, agent coalitions and their competition
Sensor fusion, feature fusion, remote monitoring, safety, reliability and fault-tolerant problems.
The Farm project provides plenty of possibilities to study and test additional examples
19.4.2023 ©L.Motus, 2004 36
Generic groups of AI applications in Real-time systems (1)
1. Automatic generation and/or assessment of alternative solutionso Typical problems – optimisation, adaptation, self-
learning, consistency check
2. Dynamic knowledge presentation and integrationo Typical problems – sensor fusion, process
monitoring and diagnosis, reliability and safety backing, pattern forming, pattern matching
19.4.2023 ©L.Motus, 2004 37
Generic groups of AI applications in Real-time systems (2)
3. Coordinated work of multiple agents (proactive components)o Typical problems – interaction of agents, multiple
goal system, dynamic change of goals, network for interacting agents
Generic groups ordered by increasing complexity :
group 1 group 2 group 3
19.4.2023 ©L.Motus, 2004 38
Characteristic issues when applying AI in Real-time systems
AI based methods cannot be applied independently and must cooperate with parts of a time-aware, or time-critical environment
Two basic goals are to be achieved – time-aware behaviour and persistent assessment the quality of service
Induction based methods create less problems time-wise, and more problems quality-wise
Deduction based methods create less problems quality-wise, and more problems time-wise
19.4.2023 ©L.Motus, 2004 39
Ways of interdisciplinary integration of AI and non-AI methods (1)
1. Mechanical combination of methods from various domainso CAD, genetic algorithms, knowledge
representation, expert systems, control theory, software engineering, qualitative reasoning
2. New methods based on combination of AI and non-AI theorieso approximate solution of hitherto not applicable
mathematical problems (Pontryagin’s maximum principle two-point boundary value problem neural nets)
19.4.2023 ©L.Motus, 2004 40
Ways of interdisciplinary integration of AI and non-AI methods (2)
3. Use of the abstract nature of AI methods to clarify the essence of problemso Intrinsic similarity of the design, analysis, and
verification of hardware and software designo Necessity to apply different methods for solving
different problems – strengths and weaknesses of algorithm-centred and interaction-centred computing
19.4.2023 ©L.Motus, 2004 41
Additional reading on Artificial Intelligence in Real-time system
Journals available in Department of Computer Control (TTU)o Engineering Applications of Artificial Intelligence
(Elsevier)o Intelligent Computer-Aided Engineering (IOC)o Real-time Systems – Journal of Time-critical Computing
(Kluwer) Other Journals
o Journal on Autonomous Agents and Multi-agent systems (Kluwer)