Introduction to Intelligent Systems PDF

Course: Intelligent Systems Introduction to Intelligent Systems Martin Molina 2020 What do we mean by intelligent system? We may see an intelligent system as a tool designed to perform tasks for us that require intelligence Examples of tasks: Specialized tasks in professional domains – – Tedious tas...

Course: Intelligent Systems Introduction to Intelligent Systems Martin Molina 2020 What do we mean by intelligent system? We may see an intelligent system as a tool designed to perform tasks for us that require intelligence Examples of tasks: Specialized tasks in professional domains – – Tedious tasks – – Medical diagnosis (e.g., recognize tumors on x-ray images) Airport gate assignment Autonomous car driving Domestic tasks (e.g., house cleaning) Dangerous tasks – Exploration of unknown areas (e.g., underwater exploration) 1 We are in an engineering context We are not interested in deciding whether a system is intelligent or not We are interested in having tools for systems engineers Tools like design metaphors, architectural patterns, computational methods and software tools Engineers who need to conceive, analyze, design and program efficiently intelligent systems 2 How can we characterize an intelligent system? We can distinguish three main properties: 1. Working in a complex world 2. Primary cognitive abilities (e.g., perception, language use, etc.) 3. Complex intelligent behavior (e.g., rationality, learning, etc.) Molina, Martin (2020). What is an intelligent system?. ArXiv preprint arXiv:2009.09083 https://arxiv.org/pdf/2009.09083.pdf 3 How can we characterize an intelligent system? We can distinguish three main properties: 1. Working in a complex world 2. Primary cognitive abilities (e.g., perception, language use, etc.) 3. Complex intelligent behavior (e.g., rationality, learning, etc.) 4 Property 1: Working in a complex world An intelligent system operates in an environment and interacts with other agents (a user or other individuals) The system observes features from the environment through sensors and performs actions using actuators The use of sensors and actuators (real or virtual) separates the body of the intelligent system from the rest of the environment (“embodiment”) Sense Sensors Sensors Intelligent system Environment Communicate Other agents Actuators Actuators Act 5 Example: Thermostat Sense [Room temperature] Communication [Manual controller] Environment [House] Performance measure [Energy consumption] Thermostat Act [Heat/not heat] 6 Other examples System Self-driving car Medical diagnosis system Chemistry tutor system Environment Roads, cars, pedestrians, … Patients Chemistry students User Observable features Actions Performance measure Passenger Images from cameras, coordinates from GPS, speed Steering, accelerator, brake, signal, horn Safety, travel time, comfort, fuel consumption Physician Test results, medical history, etc. Drug prescriptions, proposed tests Health, costs of tests and treatment Answers given by students Tests, proposed exercises, proposed readings Student’s score on tests Instructor 7 There are different properties that define the complexity of the environment with respect to the system [Russell, Norvig, 2009] Static / dynamic The environment (doesn’t change / changes) while an agent is deliberating Discrete / continuous The state of the environment, time, percepts or actions (are discrete / are continuous) Sense Fully-observable / partial-observable Sensors (detect / don’t detect) all aspects that are relevant to the choice of action Deterministic / stochastic The next state of the environment (is / isn’t) completely determined by the current state and the action Intelligent system Environment Episodic / sequential Actions (don’t influence / influence) future actions Act Known / unknown The outcomes for actions (are known / are not known) by the agent in advance Russell, S., Norvig P. (2009). Artificial Intelligence: A Modern Approach (3rd edition). Pearsons Education Limited. 8 Example: Environment of a chess player Chess player Static / dynamic Discrete / continuous Fully-observable / partial-observable Deterministic / stochastic Episodic / sequential Known / unknown 9 Example: Environment of a self-driving car Self-driving car Static / dynamic Discrete / continuous Fully-observable / partial-observable Deterministic / stochastic Episodic / sequential Known / unknown 10 How can we characterize an intelligent system? We can distinguish three main properties: 1. Working in a complex world 2. Primary cognitive abilities (e.g., perception, language use, etc.) 3. Complex intelligent behavior (e.g., rationality, learning, etc.) 11 We can identify multiple cognitive abilities 1993 Cognitive ability: Ability that requires to process mental information John B. Carroll Professor of Psychology University of Chicago (1920 -2003) Carroll, J. B. (1993). Human cognitive abilities: A survey of factor-analytic studies. Cambridge University Press. 12 We follow a pragmatic engineering approach to identify cognitive abilities We identify abilities that: – Are common in computer systems – Can be implemented with AI methods We consider two separated levels: – Primary abilities (basic abilities) – Secondary abilities (abilities that use models of other abilities) 13 Property 2: Primary cognitive abilities Other agents Reasoning about the world and making decisions about what to do Interaction Deliberation Perception Extraction of relevant data from the observed world Interaction with other agents (e.g., using language) Action control Environment Control the execution of the own actions 14 Example: Autonomous car Passenger Interaction (Passenger requests destination) Deliberation (Path planning to reach the requested destination) Perception Pedestrians Traffic signals Action control Vehicles Data extraction from ultrasonic sensors, radar, lidar, camera and GPS Steering Environment Acceleration Braking Control of driving mechanisms 15 Perception and action control are usually divided in multiple components Other agents Interaction Serial Deliberation Perception Parallel P1 P2 Action control Pm A1 A2 An Environment 16 The gap between deliberation and perception-actuation requires specific abilities Other agents Interaction Deliberation Gap Attention mechanisms Symbol grounding Execution control Perception P1 P2 Cognizant failure Action control Pm A1 A2 An Environment 17 “Action control” provides “reactive behavior” Deliberation Other agents Interaction Making decisions about what to do based on justifiable reasons Advantages: Reactive behaviors can be inhibited to reach more useful long-term goals Decisions are consistent with own knowledge Deliberation Perception Action control Environment Reactive behavior Generation of instantaneous actions in response to a stimulus (e.g., animal reflexes or in decisions based on intuitions). Advantage: Efficient reaction to dynamic events in a dynamic environment (it uses limited memory about the world) 18 We can distinguish two types of systems according to who acts in the environment Advisor system Autonomous system Acts in the environment to help the user The system makes decisions about what to do Sense Environment Helps the user to act in the environment The user makes decisions about what to do Perform task T for me Intelligent autonomous system Completion Sense User Environment Intelligent advisor system Help me perform task T Suggestion User Act Act (a) (b) 19 An intelligent system may interact with other system Sense Sense Request Intelligent system 1 Environment 1 Answer Intelligent system 2 Environment 2 Act Act 20 Intelligent systems can be part of multiagent systems creating complex organizations Communicate Communicate Agent Sense and act Partial environment Communicate Communicate Communicate Agent Sense and act Partial environment Agent Sense and act Partial environment Communicate Agent Sense and act Partial environment Agent Sense and act Partial environment Global environment 21 How can we characterize an intelligent system? We can distinguish three main properties: 1. Working in a complex world 2. Primary cognitive abilities (e.g., perception, language use, etc.) 3. Complex intelligent behavior (e.g., rationality, learning, etc.) 22 Property 3: Complex intelligent behavior Other agents Interaction Deliberation 3.1. Acting rationally 3.2. Adaptation through learning 3.3. Introspection Perception Action control Environment 23 Sub-property 3.1: Acting rationally A system acts rationally if it makes decisions to obtain the maximum performance measure Examples: A chess player selects the movement that maximizes the expectation of winning the game A self-driving car selects the best route to reach a destination considering possible traffic jams Implementation: The expected performance measure of actions is usually uncertain Rational behavior can be explicitly programmed using algorithms from decision theory (with a probabilistic representation) 24 Rational decisions affect different cognitive abilities Other agents What is the next question to ask the user?, … Interaction What is the right action to do?, what is the right method to perform a task?, … Deliberation What part of the environment requires more attention?, … Perception Action control What is the right method to control an action?, … Environment 25 Sub-property 3.2: Adaptation through learning The system is capable of improving its performance over multiple interactions with the world Examples: A chess player improves its capacity to win by learning from game experience A self-driving car reduces the time to reach destination in a city by learning from the experience of urban trips 26 Adaptation through learning can affect different cognitive abilities Other agents Learning user preferences, … Interaction Learning by deduction using a world model, … Deliberation Learning relevance of features, … Perception Action control Environment Improving action control by learning (object manipulation, …) 27 Sub-property 3.3: Introspection Capacity to analyze one's cognitive abilities – The system uses an observable model of its own abilities – This model is used to simulate self-awareness processes Practical utility: Allows the system to judge its own actions – This provides feedback to be able to learn (this feedback can also be done by simulating reactive feelings) Allows the system to generate explanations – E.g., the system is able to justify recommended decisions to the user 28 Summary of properties of an intelligent system 1. Working in a complex world – – 2. 3. Environment Other agents (e.g., user) Primary cognitive abilities – – – – Perception Action control Deliberation Interaction Complex intelligent behavior – – – Other agents Acting rationally Adaptation through learning Introspection Intelligent system Interaction Deliberation Perception Acting rationally Adaptation through learning Introspection Action control Environment 29 There are multiple sources of information about intelligent systems Scientific journals IEEE intelligent systems (IEEE) Knowlegde-based systems (Elsevier) Expert systems with applications (Elsevier) Engineering applications of artificial intelligence (Elsevier) International journal on artificial intelligence tools (World Scientific) Associations AAAI: http://aaai.org ECCAI: http://www.eccai.org 30 Lecture slides of master course “Intelligent Systems”. © 2020 Martin Molina This work is licensed under Creative Commons license CC BY-SA 4.0: https://creativecommons.org/licenses/by-sa/4.0/legalcode Suggested work citation: Molina, M. (2020): “Intelligent Systems”. Master course (lecture slides). Department of Artificial Intelligence. Universidad Politécnica de Madrid. 31