Artificial Intelligence and Expert Systems
Lecture 1
September 27, 2024
Table of Contents
1. History of Artificial Intelligence
2. What is Artificial Intelligence?
3. The Four Categories of AI
4. Differences between LLMs and General AI
5. Decision Making in AI
6. Designing Rational Agents
7. AI Environments
8. Course Topics: Extended
2/31
A Brief History of AI
• 1943: Boolean circuit model of the brain by McCulloch & Pitts.
• 1950: Turing’s seminal paper “Computing Machinery and Intelligence.”
• 1956: Official naming of the field at the Dartmouth Conference.
• 1965: Introduction of Robinson’s algorithm for logical reasoning.
• Late 1980s: Decline of expert systems leading to the first AI Winter.
• 1990s: Resurgence with statistical methods, focus on uncertainty.
• Early 2000s to now: Deep learning, big data, and AI in everyday tech.
• 2023 ChatGPT
3/31
Defining Artificial Intelligence
• AI involves machines
performing tasks that require
human intelligence.
• Encompasses learning,
reasoning, problem-solving,
perception, and language
understanding.
Figure: Image Source [Link]
24/artificial-intelligence-2/
4/31
The Four Categories of AI
• Acting Humanly
• Mimicking human behavior.
• Thinking Humanly
• Simulating human thought processes.
• Thinking Rationally
• Solving problems with logic.
• Acting Rationally
• Taking optimal actions.
5/31
AI, ML, DL ?
Figure: Image Source
6/31
LLMs vs. General AI
• LLMs: Specialized in
processing and generating
human-like text.
• i.e. ChatGPT, LlaMA, Claude,
Gemini.
• Next Word Prediction, MLM
• Performance is so far
limited Valmeekam et al.
(2024)
• “Has it all been solved ?”
Ignat et al. (2024)
• General AI: Broader, aiming to
mimic or surpass overall Figure: Image Source
human cognitive abilities. [Link]
7/31
Natural Language Processing
• Early NLP: Rule-based
systems.
• Statistical Methods: Machine
learning algorithms.
• Introduction of Transformers:
Revolutionized NLP.
• BERT and Beyond
• LLMs: Advanced dialogue
models.
• Multi-Modals (Dall-E 3)
Figure: Evolution of NLP, Image Source 8/31
Computer Vision
• 1998 - LeNet
• 2012 - AlexNet
• 2014 - VGGNet
• 2015 - ResNet
• 2015 - InceptionNet
• 2016 - YOLO
• 2017 - DenseNet
• 2023 - DALL·E 3:
• 2023 - SAM: Figure: Evolution of Computer Vision, Image Source
Segment Anything
Model.
9/31
Robotics
• First Industrial Robots: Automation in manufacturing.
• AI in Robotics: Learning and adaptation.
• Service Robots: Public and domestic use.
• Surgical Robotics: Precision in medical procedures.
• Cobots: Collaboration with humans.
• Autonomous Robots: Exploration and rescue missions.
10/31
Decision Making in AI
• AI used in multiple domains.
• Optimizing airline and military logistics.
• Automated help desks and recommendations.
• AI assists in medical diagnosis.
• Automating and optimizing decision
processes. Figure: Image Source
Figure: Image Source 11/31
What is an Agent?
Figure: Image from [Link]
agent-ai-new-era-intelligent-app-development-tarun-sharma-laffc
12/31
What is Rational Agents?
• Maximize performance measure.
• Performance measured by the agent’s impact
on the environment.
• Based on percepts and knowledge.
• Adjust to task environment.
• Example: vacuum agent cleaning strategy.
Figure: An Agent, Image Source Russell
and Norvig (2010)
13/31
What Is a Rational Agent?
• Goal: Select optimal actions.
• Equation:
M×P×K→a
• Performance Measure: Maximize M given P and K.
• Example: Vacuum agent bumps into hamster.
14/31
Agent Structure: Architecture + Program
• Agent Structure:
Agent = Architecture + Program
• Architecture: Interface with the environment.
• Program: Executes decisions.
• Bounded Rationality: - Agents may not always choose the optimal action due to
time/memory limits. - This is known as bounded rationality.
15/31
Defining Rationality
• Four key factors define rationality:
• Performance measure
• Prior knowledge of the environment
• Actions the agent can perform
• Percept sequence to date
• Rational agents maximize expected
performance based on percepts and
knowledge.
Figure: Rational Agent, Image Source Russell and
Norvig (2010)
16/31
Rationality vs Omniscience
• Rationality ≠ Perfection (omniscience).
• Rational agents maximize expected performance, not guaranteed outcomes.
• Example: Crossing a street and being hit by a meteor.
Figure: Image Generated Via Dall-E 3
17/31
Rationality and Bounded Rationality
• Rationality: choose best action given percepts.
• Bounded rationality: limited by computational resources.
• Omniscient agents know outcomes, unrealistic.
• Rationality relies on expected performance.
18/31
Environment Types: Observability and Determinism
• Observable vs. Partially Observable:
• Observable: Chess/Go, Image analysis.
• Partially Observable: Poker, Car driving, Medical diagnosis.
• Deterministic vs. Stochastic:
• Deterministic: Chess/Go, Image analysis.
• Stochastic: Poker, Car driving, Medical diagnosis.
• Episodic vs. Sequential:
• Episodic: Image classification, Spam detection.
• Sequential: Chess/Go, Poker, Car driving.
• Single vs. Multi-Agent:
• Single-Agent: Medical diagnosis, Image analysis.
• Multi-Agent: Chess/Go, Poker, Car driving.
19/31
Rationality in Different Environments
Figure: Various environments, Image Source 20/31
Examples of AI Environments
Task Observable Deterministic Episodic Static Discrete Agents
Autonomous Drones Partially Stochastic Sequential Dynamic Continuous Multi
Video Game AI Fully Stochastic Sequential Dynamic Discrete Multi
Stock Trading Partially Stochastic Sequential Dynamic Continuous Multi
Wildlife Monitoring Partially Stochastic Episodic Dynamic Continuous Single
Virtual Assistant Fully Stochastic Sequential Static Discrete Single
Weather Prediction Partially Stochastic Sequential Dynamic Continuous Single
Puzzle Solving AI Fully Deterministic Episodic Static Discrete Single
21/31
Types of Agents
• Simple reflex agents: respond to current percepts.
• Model-based agents: track internal state.
• Goal-based agents: work towards future goals.
• Utility-based agents: maximize expected utility.
22/31
Table Driven Agent
Algorithm 1 Table-Driven-Agent
1: function TABLE-DrIVEN-AGENT(percept)
2: persistent: percepts, a sequence, initially empty
3: table, a table of actions, indexed by percept sequences, initially fully
specified
4: begin
5: append percept to the end of percepts
6: action ← LOOKUP(percepts, table)
7: return action
8: end function
23/31
Simple Reflex Agent
Figure: Image Source Russell and Norvig (2010) 24/31
Model Based Reflex Agent
Figure: Image Source Russell and Norvig (2010)
25/31
Goal Based Agent
Figure: Image Source Russell and Norvig (2010) 26/31
Learning Agents
• Improve performance through
experience.
• Components: critic, learning element.
• Problem generator encourages
exploration.
• Example: taxi agent improves driving
strategy.
Figure: Image Source Russell and Norvig (2010)
27/31
Course Topics:
• Rational Agent
• Classical and Blind Search Problems
• Informed Search
• Adversarial Search
• Constraint Satisfaction
• Machine Learning (Linear Logistic Regression)
• Natural Language Processing
• Computer Vision
• Reinforcement Learning
28/31
Questions and Discussion
Any questions?
29/31
Reference
Some slides and ideas are adapted from the following sources:
• Artificial Intelligence: A Modern Approach Russell and Norvig (2010)
30/31
Thank you for your attention
September 27, 2024
Ignat, O., Jin, Z., Abzaliev, A., Biester, L., Castro, S., Deng, N., Gao, X., Gunal, A. E., He, J.,
Kazemi, A., Khalifa, M., Koh, N., Lee, A., Liu, S., Min, D. J., Mori, S., Nwatu, J. C.,
Perez-Rosas, V., Shen, S., Wang, Z., Wu, W., and Mihalcea, R. (2024). Has it all been
solved? open NLP research questions not solved by large language models. In
Calzolari, N., Kan, M.-Y., Hoste, V., Lenci, A., Sakti, S., and Xue, N., editors,
Proceedings of the 2024 Joint International Conference on Computational
Linguistics, Language Resources and Evaluation (LREC-COLING 2024), pages
8050–8094, Torino, Italia. ELRA and ICCL.
Russell, S. and Norvig, P. (2010). Artificial Intelligence: A Modern Approach. Prentice
Hall, 3 edition.
Valmeekam, K., Marquez, M., Olmo, A., Sreedharan, S., and Kambhampati, S. (2024).
Planbench: an extensible benchmark for evaluating large language models on
planning and reasoning about change. In Proceedings of the 37th International
Conference on Neural Information Processing Systems, NIPS ’23, Red Hook, NY, USA.
Curran Associates Inc.
31/31
