Artificial Intelligence Course Overview and Schedule, Lecture notes of Artificial Intelligence

Download Artificial Intelligence Course Overview and Schedule and more Lecture notes Artificial Intelligence in PDF only on Docsity!

Artificial Intelligence

CS 165A

Mar 14 , 2019

Instructor: Prof. Yu-Xiang Wang

® Review

Announcement

• HW4 due

• MP2 due

• A tournament will be conducted and the results revealed on Piazza.

• HW3 grading is being done.

• The TA catching a homework deadline.
• It will be distributed in Friday’s discussion class.
• The TA will talk about HW3 and HW4 solutions.

• Final: Next Monday 12 - 3!

MP1: Leaderboard

1. Brian Humphreys: 100%

• Multinomial NaiveBayes: Using up to 5 - grams features.

2. Claudia Zeng: 99.93%

3. Calvin Wang: 98.97%

• Other winners will be notified in private emails.

• Number 4, 5 get 20 bonus points.
• Number 6- 10 get 10 bonus points.

MP1: Considerations

• Accuracy:

• 50% is trivial.
• 80% is the Baseline multinomial naïve bayes without additional feature engineering.

• Run time: From ~10 s to more than one hour

• Use hashtables: dictionary. Use sparse representation.

• Modularize your code

• Feature extractor(Text) à Feature
• Train(Feature, Label) à Model parameters
• Predict(Feature) à Prediction
• Evaluate(Prediction, TrueLabel) à Accuracy

Final exam

• A small fraction will be about Lecture 1- 9

• Main focus on Lecture 10 – 17

• Search
• RL
• Logic

• There will be questions about using tools from 1 - 9 on the

second half of the class

Lecture 1 - 2: AI Overview

• Strong AI / Weak AI

• Turing Test

• AI for problem solving

• Rational agents

• Examples of AI in the real world

Different Ways of Looking at the AI

• Agent types / level of intelligence

• Low-level: Reflex agents
• Mid-level: Goal-based / Utility-based agents: planning
• High-level: Knowledge-based: Logic agents

• Optimization view

• Everything is an optimization problem

• Theoretical aspects

• Time/space complexity
• Algorithms and data structures
• Statistical properties: sample complexity

Rational agents and Optimization

• What is the objective function that an AI agent optimizes?

• Likelihood, Training error, Utility, Reward, Regret

• What is the argument over which the AI agent is

optimizing?

• Policy, action, search strategy

• What are the input from the environment into that

objective functions

• Observation, State, Reward, Feedback, Labels, Features.

• What are the algorithms used for these agents

• Gradient descent, SGD. Tree-search, Graph search. Dynamic programming. Explore-Exploit. 11

Example: Flu and measles

Flu

Fever Measles

Spots

P(Flu) = 0. P(Measles) = 0. P(Flu) P(Measles) P(Spots | Measles) P(Fever | Flu, Measles) P(Spots | Measles) = [0, 0.9] P(Fever | Flu, Measles) = [0.01, 0.8, 0.9, 1.0] Compute P(Flu | Fever) and P(Flu | Fever, Spots). Are they equivalent? CPTs:

d-separation and Markov Blanket

3 ways to block paths from X to Y , given E The set of nodes E d-separates sets X and Y

1. Parents
2. Children
3. Children’s other parents

Lecture 6 - 9: Machine Learning

• Types of machine learning: Supervised / unsupervised…

• Examples.

• ML is about coming up with objectives to optimize

• Often MLE, MAP.
• Often there is a graphical model.

• How to optimize?

• Gradient Descent, SGD

• Statistical Learning theory

• uniform convergence using
• Hoeffding’s inequality and Union bound.

Understanding Machine Learning

• What do we mean by saying: ML works

• Error decomposition
• Empirical risk, Risk, Generalization

• When ML does not work? What are the assumptions?

• iid: Independent and Identically Distributed
• Training data and test data are drawn from the same distribution

• Statistical Learning theory

• How many data points do we need?
• Hoeffding’s inequality and the union bound.

Example: Romania

How do we evaluate a search algorithm?

• Primary criteria to evaluate search strategies
  • Completeness
    • Is it guaranteed to find a solution (if one exists)?
  • Optimality
    • Does it find the “best” solution (if there are more than one)?
  • Time complexity
    • Number of nodes generated/expanded
    • (How long does it take to find a solution?)
  • Space complexity
    • How much memory does it require?
• Some performance measures
  • Best case
  • Worst case
  • Average case
  • Real-world case *Note that this is not saying it’s space/time complexity is optimal.