Introduction to AI/Intelligent Agents Lecture 1 PDF

Dr Josef Bajada Department of Artificial Intelligence University of Malta...

Dr Josef Bajada Department of Artificial Intelligence University of Malta [email protected] Introduction to AI / Intelligent Agents I C S 1 0 1 7 / I C S 1 0 2 0 – F O U N D AT I O N S O F A R T I F I C I A L I N T E L L I G E N C E Agenda 1. Definition of Artificial Intelligence 2. Applications of AI 3. Intelligent Agents 4. Types of AI Problems © JOSEF BAJADA 2 What is Intelligence? §Logical Reasoning §Inference of Missing Information §Understanding and Abstraction §Problem Solving §Learning from experience §Decision Making §Planning §Creativity There is no single definition of §Consciousness and Self-awareness Intelligence © JOSEF BAJADA 3 Traits of Intelligence in Humans and Animals §Process ambiguous and noisy information. §Manage lack of observability. §Handle uncertainty. §Solve problems that are too complex for conventional algorithms. §Assimilate new knowledge. §Adapt to a dynamic environment. §Compete for reward against adversaries. Can we computationally emulate some of these traits ? © JOSEF BAJADA 4 Watch how crows are capable of solving a Planning problem https://www.youtube.com/watch?v=cbSu2PXOTOc 4 mins © JOSEF BAJADA 5 What is Artificial Intelligence? “The theory and development of computer systems able to perform tasks normally requiring human intelligence, such as visual perception, speech recognition, decision-making and translation between languages” Oxford Dictionary © JOSEF BAJADA 6 A 1950s View: The Turing Test Can a computer program fool a human into thinking that the chat is with another human? Alan Turing https://www.youtube.com/watch?v=4VROUIAF2Do © JOSEF BAJADA 7 Narrow vs General AI Most advancements ARTIFICIAL GENERAL NARROW AI are in this area INTELLIGENCE (AGI) §Specialisation in a specific domain. §Apply intelligent system to any § Solve a puzzle problem. § Play a game of Chess or Go §Assimilate concepts. § Recognise Speech § Identify objects in images §Sentience. § Recognise Faces § Feel, perceive, and experience. § Correlate symptoms §Consciousness. § Control vehicles § Awareness of one’s body and surroundings. §Needs reconfiguration, or a new algorithm to solve a different task. §Self-awareness. § Awareness of consciousness and one- self. © JOSEF BAJADA 8 Main AI Research Areas Reasoning Learning Problem Perception Solving © JOSEF BAJADA 9 Applications of AI: Robotics §Industrial Robotics § Assembly Lines § Packaging and Warehouse Management §Autonomous Systems § Self-driving Cars § Drones § Space (rovers, space vehicles) §Domestic Robots § Cleaning Robots § Customer Service § Assistants / Companions (e.g. helping and monitoring the elderly) © JOSEF BAJADA 10  Applications of AI: Industrial Automation §Intelligent Control § Power Plants and Smart Grids § Manufacturing § High Risk Operations §Defect detection and prevention § Early fault detection § Quality control § Problem diagnosis § Removal of Human Error §Safety and Security § Pre-empt disaster situations. § Detect when humans are in danger. §Optimisation § Resource usage § Supply chain management § Just-in Time Systems © JOSEF BAJADA 11 Applications of AI: Health §Interpretation of Patient Data § X-Ray, MRI and CT Scan Images § EEG and ECG Wave Analysis §Genomic Diagnostics § Genetic susceptibility to diseases. § Finding genes that resist certain diseases. §Drug Design and Development §Operating Theatre Robotics §Community Data Analysis § Outbreaks of Epidemics and Pandemics § Superbug resistance § Societal Tendencies (obesity, diabetes, etc.) § Treatment efficacy © JOSEF BAJADA 12 Applications of AI: Games §Simulated Adversaries § Board games (Chess, Go, etc.) § Strategy Games § Combat AI §Non-Player Characters § Interaction with Player. § Establishing dialogue. § Interaction with simulated environment. §Virtual and Augmented Reality § Processing sensor data. § Detecting user intent. § Detection of physical environment. © JOSEF BAJADA 13 Applications of AI: Others §Personal Assistance §Education §Military / Search and Rescue §Hostile Environments (Space, Underwater, Volcanoes, Mining) §Economics and Finance (Risk, Portfolio Management, Trading) §Transportation and Logistics §Astronomy §Weather Forecasting and Climate Monitoring §Agriculture © JOSEF BAJADA 14 Watch AI beat a human F-16 pilot 5 – 0 in simulation at the DARPA AlphaDogfight Challenge https://www.youtube.com/watch?v=IOJhgC1ksNU 11 mins © JOSEF BAJADA 15 Intelligent Agents “… an autonomous entity which acts, directing its activity towards achieving goals, upon an environment using observation through sensors and consequent actuators.” Wikipedia An agent is anything that can be viewed as perceiving its environment through sensors and acting upon that environment through actuators. Russell and Norvig © JOSEF BAJADA 16 Perception-Action Cycle Intelligent Agent Sensors ? Environment Actuators Agent Function © JOSEF BAJADA 17  Percept Sequence §The Percept Sequence is the complete history of all the data the agent received from its sensors. Agent Function Percept Action Sequence Maintaining a complete history is not always feasible, and enumerating all permutations of possible percept sequences is practically impossible. © JOSEF BAJADA 18 Rational Agents Does the intelligent agent do the right thing? Rationality depends on: § The performance measure that defines the success criterion. § The agent’s prior knowledge of the environment. § The actions that the agent can perform. § The agent’s percept sequence to date. This might not be as easy as it sounds … © JOSEF BAJADA 19 Observability Not all the necessary information to make an optimal decision might be available to the agent’s sensors. These kinds of environments are called partially observable. Agent Scenario Observability A chess playing agent can see the whole board. Fully Observable A poker playing agent only has visibility of its own cards. Partially Observable An autonomous car control agent can only sense what is Partially Observable immediately coming up on the road. Maintaining historical information in memory (e.g. counting cards) often helps an agent to make better decisions under partial observability. © JOSEF BAJADA 20 Stochasticity The outcomes of an agent’s actions are not necessarily fully predictable. § An action is deterministic if the effect of applying it always results in the same outcome. §An action is stochastic if the effect of applying it can result in different outcomes. Agent Scenario Outcome A backgammon playing agent throwing a dice. Stochastic A chess playing agent making a move. Deterministic An autonomous car control agent applying the brakes. Stochastic Maintaining a probability distribution of the possible outcomes of an action often helps an agent to make better decisions under stochasticity. © JOSEF BAJADA 21 Discrete vs Continuous §An environment is discrete if there is a finite number of action choices and a finite number of states to represent it. §An environment is continuous if the space of possible states or actions that could be applied may be infinite. Agent Scenario Environment A chess game has a finite number of board positions and a Discrete finite number of possible moves. A game of tennis can be played in infinitely many Continuous combinations of states and possible actions players could make (angle, strength, player position, etc.) An autonomous car has to deal with infinitely many scenarios Continuous and the possible actions are infinite. © JOSEF BAJADA 22 Adversarial An environment is adversarial if the agent is competing against other agents (possibly humans) to achieve its objective. Agent Scenario Environment An agent playing a chess game. Adversarial An agent playing the game of poker. Adversarial An agent controlling an autonomous vehicle. Benign An agent trying to predict the weather. Benign In an adversarial scenario, the agent needs to not only consider the sensor data, but also try to predict what the opponent is likely to do. © JOSEF BAJADA 23 Simple Reflex Agent §Uses the current percept, ignoring the rest of the percept history. §Assumes the environment is fully observable. Simple Reflex Agent What does the world look like now? Sensors Environment Condition- Action Rules Select Action Actuators © JOSEF BAJADA 24  Model-Based Reflex Agent Keeps track of part of the world it can’t see by maintaining an internal state that depends on the percept history, a model of how the environment evolves, and what effects the applied actions had on the environment. Model-Based Reflex Agent State What does Sensors How the world the world evolves look like Estimate of now? What the actions partially Environment do to the world observable environment Condition-Action Rules Select Action Actuators © JOSEF BAJADA 25 Goal-Based Agent Plans ahead into the future and selects actions according to whether they eventually lead to its goals. Goal-Based Agent State What does the world look like now? Sensors How the world evolves What will the world What the actions do to the world look like after applying Environment Action A? Goals Select Action Actuators Look Ma!.... No rules! © JOSEF BAJADA 26 Utility-Based Agent Plans ahead into the future and selects actions according to whether they maximise some utility (some measure of the solution quality). Utility-Based Agent State What does the world look like now? Sensors How the world evolves What will the world What the actions look like after applying do to the world Action A? Environment Utility Function Is this state better? Can prioritise between conflicting goals and Select Action Actuators account for stochasticity. © JOSEF BAJADA 27  Learning Agent Performance Standard Learning Agent Critic Sensor Data Sensors Feedback changes Learning Element Performance Element Environment Learning knowledge Goals exploration Problem Generator Select Action Actuators © JOSEF BAJADA 28 AI to Manage Uncertainty Artificial Intelligence can be seen as the discipline of managing uncertainty through computation techniques. Uncertainty can be introduced from various sources: §Exploring all permutations is too complex or impossible. §The environment is partially observable. §The outcomes of some actions can be stochastic. §The environment is adversarial or multi-agent. §Sensor inaccuracy and noise. §Dynamic Environments. © JOSEF BAJADA 29 Coming up … §How do we solve problems with AI? §How is knowledge about the world represented? §How can we handle uncertainty? §What can an agent learn from past observations? §How does an agent perceive the world? §How does an agent understand humans? §How does an agent interact with the physical world? §Can we ensure an AI agent is ethical? © JOSEF BAJADA 30 Further Reading Stuart J. Russell and Peter Norvig, “Artificial Intelligence: A Modern Approach”, Pearson, 2016 Chapter 1 Chapter 2 David L. Poole and Alan K. Mackworth, “Artificial Intelligence: Foundations of Computational Agents”, 2nd Edition, Cambridge University Press, 2017 Chapter 1 Chapter 2 © JOSEF BAJADA 31