![Page 1: CS-E4800 Artificial Intelligence · Robotics (motion planning, sensing, ...) This Course Reasoning and problem-solving logical reasoning probabilistic reasoning Decision-making and](https://reader035.vdocument.in/reader035/viewer/2022062604/5fc3d06786ee1d071171cc8a/html5/thumbnails/1.jpg)
CS-E4800 Artificial Intelligence
Jussi Rintanen
Department of Computer ScienceAalto University
January 9, 2020
![Page 2: CS-E4800 Artificial Intelligence · Robotics (motion planning, sensing, ...) This Course Reasoning and problem-solving logical reasoning probabilistic reasoning Decision-making and](https://reader035.vdocument.in/reader035/viewer/2022062604/5fc3d06786ee1d071171cc8a/html5/thumbnails/2.jpg)
A. I. = Computerized Intelligence
Software programs that carry out tasks requiring intelligence
Robots and other autonomous devices with intelligence
Systems exhibiting intelligence
![Page 3: CS-E4800 Artificial Intelligence · Robotics (motion planning, sensing, ...) This Course Reasoning and problem-solving logical reasoning probabilistic reasoning Decision-making and](https://reader035.vdocument.in/reader035/viewer/2022062604/5fc3d06786ee1d071171cc8a/html5/thumbnails/3.jpg)
What is Intelligence?
Intelligence isAbility to act in and adapt to complex situations and circumstances.
![Page 4: CS-E4800 Artificial Intelligence · Robotics (motion planning, sensing, ...) This Course Reasoning and problem-solving logical reasoning probabilistic reasoning Decision-making and](https://reader035.vdocument.in/reader035/viewer/2022062604/5fc3d06786ee1d071171cc8a/html5/thumbnails/4.jpg)
What is Needed for Intelligence?
Core ability:
Intelligent behavior: To act intelligently
Core ability supported by:
Situational awareness (sensing: vision, hearing, ...)
Thinking, reasoning, inference
Decision-making, planning
Learning and adaptation
![Page 5: CS-E4800 Artificial Intelligence · Robotics (motion planning, sensing, ...) This Course Reasoning and problem-solving logical reasoning probabilistic reasoning Decision-making and](https://reader035.vdocument.in/reader035/viewer/2022062604/5fc3d06786ee1d071171cc8a/html5/thumbnails/5.jpg)
What is Needed for Intelligence?
Core ability:
Intelligent behavior: To act intelligently
Core ability supported by:
Situational awareness (sensing: vision, hearing, ...)
Thinking, reasoning, inference
Decision-making, planning
Learning and adaptation
![Page 6: CS-E4800 Artificial Intelligence · Robotics (motion planning, sensing, ...) This Course Reasoning and problem-solving logical reasoning probabilistic reasoning Decision-making and](https://reader035.vdocument.in/reader035/viewer/2022062604/5fc3d06786ee1d071171cc8a/html5/thumbnails/6.jpg)
Why Is A.I. Studied?
Basic research questions:
What IS intelligence?
What are the computational mechanisms behind intelligence?
How to create computation (software) that exhibits intelligence?
![Page 7: CS-E4800 Artificial Intelligence · Robotics (motion planning, sensing, ...) This Course Reasoning and problem-solving logical reasoning probabilistic reasoning Decision-making and](https://reader035.vdocument.in/reader035/viewer/2022062604/5fc3d06786ee1d071171cc8a/html5/thumbnails/7.jpg)
Why Is A.I. Studied?
A.I. as an engineering enterprise:Automate tasks that require intelligence
intelligent softwareintelligent devices, vehiclesintelligent systems (networks/teams of devices, vehicles, ...)
What kind of software technology is needed?
What are the limits of these SW technologies?
![Page 8: CS-E4800 Artificial Intelligence · Robotics (motion planning, sensing, ...) This Course Reasoning and problem-solving logical reasoning probabilistic reasoning Decision-making and](https://reader035.vdocument.in/reader035/viewer/2022062604/5fc3d06786ee1d071171cc8a/html5/thumbnails/8.jpg)
Two Alternative Research Programs
1 Create intelligence similar to human intelligence (Strong A.I.)Imitate human cognitionImitate human behavior, skills, capabilities
2 Create SW for complex tasks requiring intelligence (Weak A.I.)No need to imitate human intelligenceCapabilities in some tasks below humansCapabilities in other tasks at human or super-human level
Observation:1 No-one knows how to do Strong AI (machines that really think) !!!2 Weak A.I. is what can be had now, and this may be more useful!
![Page 9: CS-E4800 Artificial Intelligence · Robotics (motion planning, sensing, ...) This Course Reasoning and problem-solving logical reasoning probabilistic reasoning Decision-making and](https://reader035.vdocument.in/reader035/viewer/2022062604/5fc3d06786ee1d071171cc8a/html5/thumbnails/9.jpg)
Two Alternative Research Programs
1 Create intelligence similar to human intelligence (Strong A.I.)Imitate human cognitionImitate human behavior, skills, capabilities
2 Create SW for complex tasks requiring intelligence (Weak A.I.)No need to imitate human intelligenceCapabilities in some tasks below humansCapabilities in other tasks at human or super-human level
Observation:1 No-one knows how to do Strong AI (machines that really think) !!!2 Weak A.I. is what can be had now, and this may be more useful!
![Page 10: CS-E4800 Artificial Intelligence · Robotics (motion planning, sensing, ...) This Course Reasoning and problem-solving logical reasoning probabilistic reasoning Decision-making and](https://reader035.vdocument.in/reader035/viewer/2022062604/5fc3d06786ee1d071171cc8a/html5/thumbnails/10.jpg)
Two Alternative Research Programs
1 Create intelligence similar to human intelligence (Strong A.I.)Imitate human cognitionImitate human behavior, skills, capabilities
2 Create SW for complex tasks requiring intelligence (Weak A.I.)No need to imitate human intelligenceCapabilities in some tasks below humansCapabilities in other tasks at human or super-human level
Observation:1 No-one knows how to do Strong AI (machines that really think) !!!2 Weak A.I. is what can be had now, and this may be more useful!
![Page 11: CS-E4800 Artificial Intelligence · Robotics (motion planning, sensing, ...) This Course Reasoning and problem-solving logical reasoning probabilistic reasoning Decision-making and](https://reader035.vdocument.in/reader035/viewer/2022062604/5fc3d06786ee1d071171cc8a/html5/thumbnails/11.jpg)
Why Is A.I. Important: Industrial Applications
Management of energy, utility infrastructure (power, water),communications, manufacturing, transportation & logistics,service industriesMore A.I. needed to assist and replace humans in difficult tasks:
High-level supervision: medium/long term planningExceptional situations: faults, failures, ...
Non-intelligent automation requires human supervision
Human supervision is expensive, unreliable, slow: need for A.I.
![Page 12: CS-E4800 Artificial Intelligence · Robotics (motion planning, sensing, ...) This Course Reasoning and problem-solving logical reasoning probabilistic reasoning Decision-making and](https://reader035.vdocument.in/reader035/viewer/2022062604/5fc3d06786ee1d071171cc8a/html5/thumbnails/12.jpg)
Why Is A.I. Important: Industrial Applications
Management of energy, utility infrastructure (power, water),communications, manufacturing, transportation & logistics,service industriesMore A.I. needed to assist and replace humans in difficult tasks:
High-level supervision: medium/long term planningExceptional situations: faults, failures, ...
Non-intelligent automation requires human supervision
Human supervision is expensive, unreliable, slow: need for A.I.
![Page 13: CS-E4800 Artificial Intelligence · Robotics (motion planning, sensing, ...) This Course Reasoning and problem-solving logical reasoning probabilistic reasoning Decision-making and](https://reader035.vdocument.in/reader035/viewer/2022062604/5fc3d06786ee1d071171cc8a/html5/thumbnails/13.jpg)
Why Is A.I. Important: Industrial Applications
Management of energy, utility infrastructure (power, water),communications, manufacturing, transportation & logistics,service industriesMore A.I. needed to assist and replace humans in difficult tasks:
High-level supervision: medium/long term planningExceptional situations: faults, failures, ...
Non-intelligent automation requires human supervision
Human supervision is expensive, unreliable, slow: need for A.I.
![Page 14: CS-E4800 Artificial Intelligence · Robotics (motion planning, sensing, ...) This Course Reasoning and problem-solving logical reasoning probabilistic reasoning Decision-making and](https://reader035.vdocument.in/reader035/viewer/2022062604/5fc3d06786ee1d071171cc8a/html5/thumbnails/14.jpg)
Why Is A.I. Important: SW & IT
Different types of A.I. will solve core problems of SW & IT:
Produce conventional SWcheaplyreliablyquickly
Produce complex SWComplex inputs and outputs (“Artificial Intelligence”)Optimization of SW behaviorAutomate things that were never automated before
This course covers many of the basic technologies behind these!
![Page 15: CS-E4800 Artificial Intelligence · Robotics (motion planning, sensing, ...) This Course Reasoning and problem-solving logical reasoning probabilistic reasoning Decision-making and](https://reader035.vdocument.in/reader035/viewer/2022062604/5fc3d06786ee1d071171cc8a/html5/thumbnails/15.jpg)
Why Is A.I. Important: SW & IT
Different types of A.I. will solve core problems of SW & IT:
Produce conventional SWcheaplyreliablyquickly
Produce complex SWComplex inputs and outputs (“Artificial Intelligence”)Optimization of SW behaviorAutomate things that were never automated before
This course covers many of the basic technologies behind these!
![Page 16: CS-E4800 Artificial Intelligence · Robotics (motion planning, sensing, ...) This Course Reasoning and problem-solving logical reasoning probabilistic reasoning Decision-making and](https://reader035.vdocument.in/reader035/viewer/2022062604/5fc3d06786ee1d071171cc8a/html5/thumbnails/16.jpg)
Why Is A.I. Important: SW & IT
Different types of A.I. will solve core problems of SW & IT:
Produce conventional SWcheaplyreliablyquickly
Produce complex SWComplex inputs and outputs (“Artificial Intelligence”)Optimization of SW behaviorAutomate things that were never automated before
This course covers many of the basic technologies behind these!
![Page 17: CS-E4800 Artificial Intelligence · Robotics (motion planning, sensing, ...) This Course Reasoning and problem-solving logical reasoning probabilistic reasoning Decision-making and](https://reader035.vdocument.in/reader035/viewer/2022062604/5fc3d06786ee1d071171cc8a/html5/thumbnails/17.jpg)
SW Production: From 1940ies Until 1960ies
machine code
Fortran, Cobol, C, ... UI tools DBMS
high-level specification languages
![Page 18: CS-E4800 Artificial Intelligence · Robotics (motion planning, sensing, ...) This Course Reasoning and problem-solving logical reasoning probabilistic reasoning Decision-making and](https://reader035.vdocument.in/reader035/viewer/2022062604/5fc3d06786ee1d071171cc8a/html5/thumbnails/18.jpg)
SW Production: From 1960ies Until 1970ies
machine code
Fortran, Cobol, C, ... UI tools DBMS
high-level specification languages
![Page 19: CS-E4800 Artificial Intelligence · Robotics (motion planning, sensing, ...) This Course Reasoning and problem-solving logical reasoning probabilistic reasoning Decision-making and](https://reader035.vdocument.in/reader035/viewer/2022062604/5fc3d06786ee1d071171cc8a/html5/thumbnails/19.jpg)
SW Production: From 1980ies Until 2020s
machine code
Fortran, Cobol, C, ...Fortran, Cobol, C, Java, ... UI tools DBMS
high-level specification languages
![Page 20: CS-E4800 Artificial Intelligence · Robotics (motion planning, sensing, ...) This Course Reasoning and problem-solving logical reasoning probabilistic reasoning Decision-making and](https://reader035.vdocument.in/reader035/viewer/2022062604/5fc3d06786ee1d071171cc8a/html5/thumbnails/20.jpg)
SW Production: Future
machine code
Fortran, Cobol, C, ...Fortran, Cobol, C, Java, ... UI tools DBMS
high-level specification languages
![Page 21: CS-E4800 Artificial Intelligence · Robotics (motion planning, sensing, ...) This Course Reasoning and problem-solving logical reasoning probabilistic reasoning Decision-making and](https://reader035.vdocument.in/reader035/viewer/2022062604/5fc3d06786ee1d071171cc8a/html5/thumbnails/21.jpg)
This Course
Focus of the course is in symbolic A.I.decision-making, problem-solving, planning, reasoning, inferencetechnologies needed for industrial applications
Other courses at Aalto:Machine Learning (classification, image processing, statistical NLP, ...)Robotics (motion planning, sensing, ...)
![Page 22: CS-E4800 Artificial Intelligence · Robotics (motion planning, sensing, ...) This Course Reasoning and problem-solving logical reasoning probabilistic reasoning Decision-making and](https://reader035.vdocument.in/reader035/viewer/2022062604/5fc3d06786ee1d071171cc8a/html5/thumbnails/22.jpg)
This Course
Reasoning and problem-solvinglogical reasoningprobabilistic reasoning
Decision-making and planningsequential problem-solving by searchnondeterminism, uncertainty, observabilitygame-playing and decision-making in presence of other agents
Modeling and learningadaptation and learning in decision-makingmodel-based A.I.
![Page 23: CS-E4800 Artificial Intelligence · Robotics (motion planning, sensing, ...) This Course Reasoning and problem-solving logical reasoning probabilistic reasoning Decision-making and](https://reader035.vdocument.in/reader035/viewer/2022062604/5fc3d06786ee1d071171cc8a/html5/thumbnails/23.jpg)
Symbolic vs. Sub-Symbolic A.I.
conscious / symbolic:Reasoning: Infer things from facts and observations
Decision-making: What to do to achieve ones goals?Learning: Form new undertanding from experience
sub-conscious / sub-symbolic:Vision: recognize objects, relations between them
Hearing: recognize words and language, recognize physicalevents
Tactile senses: understand physical objects and events throughtouching
![Page 24: CS-E4800 Artificial Intelligence · Robotics (motion planning, sensing, ...) This Course Reasoning and problem-solving logical reasoning probabilistic reasoning Decision-making and](https://reader035.vdocument.in/reader035/viewer/2022062604/5fc3d06786ee1d071171cc8a/html5/thumbnails/24.jpg)
Symbolic A.I.Dominant approach until 1990ies
data and knowledge representation: semantic networks, logic, ...logical reasoning, probabilistic reasoningalgorithmic decision-making and planningrule-based systems, expert systems...
Features:Explicit representations: rules, facts, diagrammatic representationsIntelligence as (complex) computation
Strengths:conscious intelligenceabstract thinking, reasoningplanning, decision-making, scheduling, ...intelligence within man-made systems = industrial applications
![Page 25: CS-E4800 Artificial Intelligence · Robotics (motion planning, sensing, ...) This Course Reasoning and problem-solving logical reasoning probabilistic reasoning Decision-making and](https://reader035.vdocument.in/reader035/viewer/2022062604/5fc3d06786ee1d071171cc8a/html5/thumbnails/25.jpg)
Symbolic A.I.Dominant approach until 1990ies
data and knowledge representation: semantic networks, logic, ...logical reasoning, probabilistic reasoningalgorithmic decision-making and planningrule-based systems, expert systems...
Features:Explicit representations: rules, facts, diagrammatic representationsIntelligence as (complex) computation
Strengths:conscious intelligenceabstract thinking, reasoningplanning, decision-making, scheduling, ...intelligence within man-made systems = industrial applications
![Page 26: CS-E4800 Artificial Intelligence · Robotics (motion planning, sensing, ...) This Course Reasoning and problem-solving logical reasoning probabilistic reasoning Decision-making and](https://reader035.vdocument.in/reader035/viewer/2022062604/5fc3d06786ee1d071171cc8a/html5/thumbnails/26.jpg)
Symbolic A.I.Dominant approach until 1990ies
data and knowledge representation: semantic networks, logic, ...logical reasoning, probabilistic reasoningalgorithmic decision-making and planningrule-based systems, expert systems...
Features:Explicit representations: rules, facts, diagrammatic representationsIntelligence as (complex) computation
Strengths:conscious intelligenceabstract thinking, reasoningplanning, decision-making, scheduling, ...intelligence within man-made systems = industrial applications
![Page 27: CS-E4800 Artificial Intelligence · Robotics (motion planning, sensing, ...) This Course Reasoning and problem-solving logical reasoning probabilistic reasoning Decision-making and](https://reader035.vdocument.in/reader035/viewer/2022062604/5fc3d06786ee1d071171cc8a/html5/thumbnails/27.jpg)
Sub-symbolic A.I.
Steadily increased in importance since 1990iesstatistical methods, especially in natural language processing (NLP)machine learning (especially connectionism, neural networks)
Premise: World is too complex to model explicitly by rules, facts
Approach: Intelligence without representation, e.g. by “learning”Strengths:
sub-conscious skills: vision, speech, ...human-computer interaction (human thinking cannot be modelled...)intelligence with and for natural systems (life sciences, medicine, ...)
![Page 28: CS-E4800 Artificial Intelligence · Robotics (motion planning, sensing, ...) This Course Reasoning and problem-solving logical reasoning probabilistic reasoning Decision-making and](https://reader035.vdocument.in/reader035/viewer/2022062604/5fc3d06786ee1d071171cc8a/html5/thumbnails/28.jpg)
Sub-symbolic A.I.
Steadily increased in importance since 1990iesstatistical methods, especially in natural language processing (NLP)machine learning (especially connectionism, neural networks)
Premise: World is too complex to model explicitly by rules, facts
Approach: Intelligence without representation, e.g. by “learning”Strengths:
sub-conscious skills: vision, speech, ...human-computer interaction (human thinking cannot be modelled...)intelligence with and for natural systems (life sciences, medicine, ...)
![Page 29: CS-E4800 Artificial Intelligence · Robotics (motion planning, sensing, ...) This Course Reasoning and problem-solving logical reasoning probabilistic reasoning Decision-making and](https://reader035.vdocument.in/reader035/viewer/2022062604/5fc3d06786ee1d071171cc8a/html5/thumbnails/29.jpg)
Sub-symbolic A.I.
Steadily increased in importance since 1990iesstatistical methods, especially in natural language processing (NLP)machine learning (especially connectionism, neural networks)
Premise: World is too complex to model explicitly by rules, facts
Approach: Intelligence without representation, e.g. by “learning”Strengths:
sub-conscious skills: vision, speech, ...human-computer interaction (human thinking cannot be modelled...)intelligence with and for natural systems (life sciences, medicine, ...)
![Page 30: CS-E4800 Artificial Intelligence · Robotics (motion planning, sensing, ...) This Course Reasoning and problem-solving logical reasoning probabilistic reasoning Decision-making and](https://reader035.vdocument.in/reader035/viewer/2022062604/5fc3d06786ee1d071171cc8a/html5/thumbnails/30.jpg)
Symbolic A.I.
“Top-down approach” to A.I.
Domain: problems that can be formalized
+ Understandable
+ Debuggable
+ Extension of conventional SW technology
– Unsuitable to poorly understood problems (e.g. speech, imagerecognition)
![Page 31: CS-E4800 Artificial Intelligence · Robotics (motion planning, sensing, ...) This Course Reasoning and problem-solving logical reasoning probabilistic reasoning Decision-making and](https://reader035.vdocument.in/reader035/viewer/2022062604/5fc3d06786ee1d071171cc8a/html5/thumbnails/31.jpg)
Symbolic A.I.
“Top-down approach” to A.I.
Domain: problems that can be formalized
+ Understandable
+ Debuggable
+ Extension of conventional SW technology
– Unsuitable to poorly understood problems (e.g. speech, imagerecognition)
![Page 32: CS-E4800 Artificial Intelligence · Robotics (motion planning, sensing, ...) This Course Reasoning and problem-solving logical reasoning probabilistic reasoning Decision-making and](https://reader035.vdocument.in/reader035/viewer/2022062604/5fc3d06786ee1d071171cc8a/html5/thumbnails/32.jpg)
Sub-Symbolic A.I.
“Bottom-up approach” to A.I.
Domain: problems that cannot be formalized
+ Basic scheme straightforward: teach, teach, teach, then deploy
+ No need to understand the domain well
– Hard to debug
– Quality guarantees only by testing
– Risky in safety-critical domains (e.g. self-driving cars!)
![Page 33: CS-E4800 Artificial Intelligence · Robotics (motion planning, sensing, ...) This Course Reasoning and problem-solving logical reasoning probabilistic reasoning Decision-making and](https://reader035.vdocument.in/reader035/viewer/2022062604/5fc3d06786ee1d071171cc8a/html5/thumbnails/33.jpg)
Sub-Symbolic A.I.
“Bottom-up approach” to A.I.
Domain: problems that cannot be formalized
+ Basic scheme straightforward: teach, teach, teach, then deploy
+ No need to understand the domain well
– Hard to debug
– Quality guarantees only by testing
– Risky in safety-critical domains (e.g. self-driving cars!)
![Page 34: CS-E4800 Artificial Intelligence · Robotics (motion planning, sensing, ...) This Course Reasoning and problem-solving logical reasoning probabilistic reasoning Decision-making and](https://reader035.vdocument.in/reader035/viewer/2022062604/5fc3d06786ee1d071171cc8a/html5/thumbnails/34.jpg)
Example: Autonomous vehicles
lots of symbolic (often algorithmic) problemsroute planningvehicle trajectories: planning, predictionman-made infrastructure (roads, junctions, parking)traffic rules
some sub-symbolic problems:anticipation of actions by non-vehicle entities (people, animals, ...) andother vehicles
Status: Not a solved problem, by far!
![Page 35: CS-E4800 Artificial Intelligence · Robotics (motion planning, sensing, ...) This Course Reasoning and problem-solving logical reasoning probabilistic reasoning Decision-making and](https://reader035.vdocument.in/reader035/viewer/2022062604/5fc3d06786ee1d071171cc8a/html5/thumbnails/35.jpg)
Example: Speech Recognition
−→ HELLO!
Mapping an audio stream to a text stream, recognizing all wordscorrectly
Difficulty: phonemes poorly defined, vary dialect to dialect, regionto region, speaker to speaker
State of the art: neural networks
Status: good enough to be practically useful
![Page 36: CS-E4800 Artificial Intelligence · Robotics (motion planning, sensing, ...) This Course Reasoning and problem-solving logical reasoning probabilistic reasoning Decision-making and](https://reader035.vdocument.in/reader035/viewer/2022062604/5fc3d06786ee1d071171cc8a/html5/thumbnails/36.jpg)
Example: Image Classification
![Page 37: CS-E4800 Artificial Intelligence · Robotics (motion planning, sensing, ...) This Course Reasoning and problem-solving logical reasoning probabilistic reasoning Decision-making and](https://reader035.vdocument.in/reader035/viewer/2022062604/5fc3d06786ee1d071171cc8a/html5/thumbnails/37.jpg)
Example: Image Classification
Recognize what an image represents
Difficulty: few things have a “dictionary definition”; What is adog? What is a cat? How do they look like?
State of the art: neural networks with supervised learning
Status: Not a solved problemRequires lots of training; limited accuracy, reliability, ability to handle high
number of categories
![Page 38: CS-E4800 Artificial Intelligence · Robotics (motion planning, sensing, ...) This Course Reasoning and problem-solving logical reasoning probabilistic reasoning Decision-making and](https://reader035.vdocument.in/reader035/viewer/2022062604/5fc3d06786ee1d071171cc8a/html5/thumbnails/38.jpg)
Example: Game of Go
Difficulty: valuing board configurations
state of the art: game-tree search, withboard configurations evaluated with aneural network
Best Go programs (AlphaGo, Alpha Zero, Zen) “symbolic” except forboard evaluation
Status: Best Go programs beat best human players
![Page 39: CS-E4800 Artificial Intelligence · Robotics (motion planning, sensing, ...) This Course Reasoning and problem-solving logical reasoning probabilistic reasoning Decision-making and](https://reader035.vdocument.in/reader035/viewer/2022062604/5fc3d06786ee1d071171cc8a/html5/thumbnails/39.jpg)
Example: Natural LanguageIdealized natural language easy:
Mary’s daughter has a dog whose name is Spark. Ann is Mary’s sister. Spark’sfather is Spot. Ann’s father is Bill.
Actual text (NY Times, Jan 2, 2017) far more difficult:
The cascade of explosions on Sept. 1 that destroyed a Falcon 9 rocket on thelaunchpad at Cape Canaveral Air Force Station in Florida was perplexing andconcerning, because it occurred during what is usually regarded as a saferportion of operations the fueling of propellants about eight minutes before theignition of the engines for a planned test. (The launch had been scheduled fortwo days later.)
Status: NLP systems today only very superficially understand language
![Page 40: CS-E4800 Artificial Intelligence · Robotics (motion planning, sensing, ...) This Course Reasoning and problem-solving logical reasoning probabilistic reasoning Decision-making and](https://reader035.vdocument.in/reader035/viewer/2022062604/5fc3d06786ee1d071171cc8a/html5/thumbnails/40.jpg)
Human-like A.I.?
Needed: Interaction with humans, replicating human skillsSurveillance A.I. understanding what humans are doingRobot communicating with a non-technical person
Not so much: artificial environmentsbuilt infrastructure: manufacturing, transportation & logisticsvirtual environments (software, internet, ...)
![Page 41: CS-E4800 Artificial Intelligence · Robotics (motion planning, sensing, ...) This Course Reasoning and problem-solving logical reasoning probabilistic reasoning Decision-making and](https://reader035.vdocument.in/reader035/viewer/2022062604/5fc3d06786ee1d071171cc8a/html5/thumbnails/41.jpg)
Non-human-like A.I.?
Humans notoriously bad withreasoning with probabilitiesreasoning with numberslogical reasoningscheduling, planning, ...
Much of value of A.I. (and computers) in things humans do badly!
If you can do better than humans, why imitate?
![Page 42: CS-E4800 Artificial Intelligence · Robotics (motion planning, sensing, ...) This Course Reasoning and problem-solving logical reasoning probabilistic reasoning Decision-making and](https://reader035.vdocument.in/reader035/viewer/2022062604/5fc3d06786ee1d071171cc8a/html5/thumbnails/42.jpg)
Intelligent SystemsInteraction with the environment
perception (sensing)
decision-making, reasoning, learning
action (actuators)
Agent Environment
Perception
Action
Sensors: camera, microphone, Internet, . . .
Actuators: wheels, arms, valves, Internet, . . .
![Page 43: CS-E4800 Artificial Intelligence · Robotics (motion planning, sensing, ...) This Course Reasoning and problem-solving logical reasoning probabilistic reasoning Decision-making and](https://reader035.vdocument.in/reader035/viewer/2022062604/5fc3d06786ee1d071171cc8a/html5/thumbnails/43.jpg)
Model-Based A.I.
Agent Environment
Perception
Action
Agent Environment
Perception
Action
Model
Planning
Important assumption: Model availableWhere does the model come from? Built manually,semi-automatically, automatically, learned, ...
![Page 44: CS-E4800 Artificial Intelligence · Robotics (motion planning, sensing, ...) This Course Reasoning and problem-solving logical reasoning probabilistic reasoning Decision-making and](https://reader035.vdocument.in/reader035/viewer/2022062604/5fc3d06786ee1d071171cc8a/html5/thumbnails/44.jpg)
General Solutions to A.I.
Partially Observable Markov Decision Processes (POMDP)Decision-making under a very general world modelComputationally extremely hard: applicable to small problems only
Reinforcement Learning (RL)Observing and acting interleaved with learningLearning rate too slow for general application
Optimal Universal A.I. (Hutter 2001)Optimal learning and decision-making (a form of RL)Even with approximation not scalable past few states
Machine LearningIn popular press Neural Networks = A.I.Solves some (narrow) problem classes, primarily classification
![Page 45: CS-E4800 Artificial Intelligence · Robotics (motion planning, sensing, ...) This Course Reasoning and problem-solving logical reasoning probabilistic reasoning Decision-making and](https://reader035.vdocument.in/reader035/viewer/2022062604/5fc3d06786ee1d071171cc8a/html5/thumbnails/45.jpg)
Real-World A.I. Systems
Conventional (software) engineering approach:
Analyze problem: What features does it have?Decompose problem:
hierarchy of problems and subproblemsinterfaces and inter-component communication
Implement solutions to all subproblems
Integrate components to a working system
![Page 46: CS-E4800 Artificial Intelligence · Robotics (motion planning, sensing, ...) This Course Reasoning and problem-solving logical reasoning probabilistic reasoning Decision-making and](https://reader035.vdocument.in/reader035/viewer/2022062604/5fc3d06786ee1d071171cc8a/html5/thumbnails/46.jpg)
General Architecture
long-term planning
medium-term planning
short-term control
sensor fusion
actuatorssensors
model /database
environment
![Page 47: CS-E4800 Artificial Intelligence · Robotics (motion planning, sensing, ...) This Course Reasoning and problem-solving logical reasoning probabilistic reasoning Decision-making and](https://reader035.vdocument.in/reader035/viewer/2022062604/5fc3d06786ee1d071171cc8a/html5/thumbnails/47.jpg)
Example: Self-Driving Cars (CMU)
![Page 48: CS-E4800 Artificial Intelligence · Robotics (motion planning, sensing, ...) This Course Reasoning and problem-solving logical reasoning probabilistic reasoning Decision-making and](https://reader035.vdocument.in/reader035/viewer/2022062604/5fc3d06786ee1d071171cc8a/html5/thumbnails/48.jpg)
Example: Self-Driving Cab Company
acquire, maintain and sell cars
assign idle cabs to towns/suburbs
assign cab to a client
InternetInternet
model /database
![Page 49: CS-E4800 Artificial Intelligence · Robotics (motion planning, sensing, ...) This Course Reasoning and problem-solving logical reasoning probabilistic reasoning Decision-making and](https://reader035.vdocument.in/reader035/viewer/2022062604/5fc3d06786ee1d071171cc8a/html5/thumbnails/49.jpg)
Examples: Chess, Checkers
Deep Blue in 1997 capable of beating best human players:endgame databases, deep non-exhaustive search of the game treehttps://www.youtube.com/watch?v=NJarxpYyoFI
Checkers is a solved game: full endgame database for max. 8pieces, exhaustive search of the game tree (program: Chinook)Schaeffer et al., Checkers is Solved, Science, 2007.
![Page 50: CS-E4800 Artificial Intelligence · Robotics (motion planning, sensing, ...) This Course Reasoning and problem-solving logical reasoning probabilistic reasoning Decision-making and](https://reader035.vdocument.in/reader035/viewer/2022062604/5fc3d06786ee1d071171cc8a/html5/thumbnails/50.jpg)
Examples: Poker
Variants of Poker played close to optimal and better than humans
Bowling et al., Heads-up Limit Hold’Em Poker is Solved, Science, 2015.
For competitive game-play, need to evaluate a large fraction of 3.19×1014
situations that could be encountered during play (current hand, cardsseen so far)Bowling et al. did this with a 2+ month computation with a cluster of200× 24 cores and 10.9 TB of disk.
Liberatus (Sandholm 2017)Plays no-limit Texas hold’em poker at super-human level, beating leadinghuman poker playersBased on sophisticated algorithms for solving game-theoretical problemsLarge-scale “learning” to assign value to different game situations
![Page 51: CS-E4800 Artificial Intelligence · Robotics (motion planning, sensing, ...) This Course Reasoning and problem-solving logical reasoning probabilistic reasoning Decision-making and](https://reader035.vdocument.in/reader035/viewer/2022062604/5fc3d06786ee1d071171cc8a/html5/thumbnails/51.jpg)
Examples: IBM Watsonhttps://www.youtube.com/watch?v=WFR3lOm_xhE
Plays the TV quiz “Jeopardy” successfully
Analysis of quiz questions, decomposition to text queries
Find matching sentences from text databases
Multiple decomposition and evaluation methods
Grades answers and offers a solution when grade is high enoughFerrucci et al., Building Watson: An Overview of the DeepQAProject, AI Magazine, 2010.
![Page 52: CS-E4800 Artificial Intelligence · Robotics (motion planning, sensing, ...) This Course Reasoning and problem-solving logical reasoning probabilistic reasoning Decision-making and](https://reader035.vdocument.in/reader035/viewer/2022062604/5fc3d06786ee1d071171cc8a/html5/thumbnails/52.jpg)
Examples: Robots in Natural Environment
Boston Dynamics: Big Doghttps://www.youtube.com/watch?v=cNZPRsrwumQ
![Page 53: CS-E4800 Artificial Intelligence · Robotics (motion planning, sensing, ...) This Course Reasoning and problem-solving logical reasoning probabilistic reasoning Decision-making and](https://reader035.vdocument.in/reader035/viewer/2022062604/5fc3d06786ee1d071171cc8a/html5/thumbnails/53.jpg)
Example: Self-Driving Cars (Waymo)
https://www.youtube.com/watch?v=TsaES--OTzM
![Page 54: CS-E4800 Artificial Intelligence · Robotics (motion planning, sensing, ...) This Course Reasoning and problem-solving logical reasoning probabilistic reasoning Decision-making and](https://reader035.vdocument.in/reader035/viewer/2022062604/5fc3d06786ee1d071171cc8a/html5/thumbnails/54.jpg)
Features of Successful AI Systems
Task-specific
Data and computation intensive (many CPUs or GPUs, largememories), rather than having “real” intelligence
Critical: detailed models, complex algorithms, and/or massivelearning
![Page 55: CS-E4800 Artificial Intelligence · Robotics (motion planning, sensing, ...) This Course Reasoning and problem-solving logical reasoning probabilistic reasoning Decision-making and](https://reader035.vdocument.in/reader035/viewer/2022062604/5fc3d06786ee1d071171cc8a/html5/thumbnails/55.jpg)
Why Is A.I. Difficult?
1 Difficulty of modeling the world, intelligenceHumans understand themselves poorly!examples: natural language, human concepts
2 Difficulty of computing the desired resultscomputational complexity (NP-hardness, undecidability, ...)examples: decision-making, robot motion-planning, ...
![Page 56: CS-E4800 Artificial Intelligence · Robotics (motion planning, sensing, ...) This Course Reasoning and problem-solving logical reasoning probabilistic reasoning Decision-making and](https://reader035.vdocument.in/reader035/viewer/2022062604/5fc3d06786ee1d071171cc8a/html5/thumbnails/56.jpg)
First Programming AssignmentTurn this to this:
Human sees the trivial solution of moving each piece once
Computer sees an astronomic number of possibilities
![Page 57: CS-E4800 Artificial Intelligence · Robotics (motion planning, sensing, ...) This Course Reasoning and problem-solving logical reasoning probabilistic reasoning Decision-making and](https://reader035.vdocument.in/reader035/viewer/2022062604/5fc3d06786ee1d071171cc8a/html5/thumbnails/57.jpg)
First Programming Assignment
Your task, given a code template and a search algorithm:1 Implement Chess board with one or more knights and their moves2 Experiment moving groups of knights from initial to goal positions
Sample code implements a basic breadth-first search algorithm:has difficulty solving these problems with more than 4 or 5 knights
Next week we will see other, far better, methods