Friday, November 4, 2016

Intro to AI

finish 40% in 3 days, may come back if needed.

1 AI in general

  • Finace: making trading decisions
  • Web
  • Game
  • Robotics
  • medicine: diagnostic
  • languague translation


  • Fully vs. Partially Observable
  • Deterministic vs. Stochastic
  • Discrete vs. Continuous
  • benign vs. Adversarial environment
AI as uncertainty management:
  • sensor limit
  • ignorance
  • stochastic

definition of a problem

  • initial state
  • actions(s)
  • result(s, a): return new state
  • GoalTest(s): return True/False
  • Path cost/step cost

2 finding the best route

Breath first search vs Depth first search

3 Beyes Network

complementary, independence,
conditional probability
joint probability
Bayes Rule
A is not directly observable, truth or fact of a cancer, B is directly obserbale, testing result of cancer. In Beyes Network (A—>B), It takes 1+2 parameters to specify. The core calculation depends on joint probability.
Posterior= Likelihood*Prioir/Marginal likelihood
Away effect:
indepedence does not imply conditional independence
e.g. S and R are independent, they are both necessary conditions to H. If we know H, then the existence of S or R will affect each other.
Screen Shot 2016-11-03 at 5.28.36 PM

4 Probabilistic inference

  • Evidence
  • Hidden
  • Query
variable elimination: joining table, sum out, rejection sampling, likelihood weighting.
Monty Hall problem: 1/3 vs 2/3

5 Machine learning

  • Bayes networks: reason with known models
  • machine learning: learn models from data
Stanley self-driving car: to see 200 meters far is a key factor to win the 2005 DAPRA race.
  • Classification: output is binary or a fixed number of classes
  • regression: output is continuous
Overfitting is a major source of poor performance of a machine learning algorithm.
spam detection: maximum likelihood, laplace smoothing.
Naive Bayes model, but not powerful enough to counter spam.d
logistic regression is slightly more complex than linear regression
gradient decent
perceptron algorithm
support vector machine: kernel trick
K nearest neighbors: nonparameter

6 Unsupervised learning

7 Representation with Logic

Agent design: reflex-based, goal-based, utility-based
complexity of environment
representation: logic
propositional logic:
  • V: or ;
  • ^: and
  • can’t handle uncertainty
  • can’t handle objects
  • no short cuts
first-order logic: relation to object

8 planning

problem solving: in deterministic and fully obserable environment, have starting point and end goal, find a path.
In your real life, you do your planning ahead of time, you have the complete goal, then without interacting with the world, without sensing it at all, start execute the path.
However, if people have gone out and blindfolk walk, they can never walk in a straight line, without any feedback from the environment.
In a stochastic environment, we must react to unexpected result. If the road is closed, we don’t know our state, which have partial observability.

19 Robotics

Robotics is the science of bridging the gap between sensor data and actions.
Kinematic state of a car: 3D of state space: xy plane + steer direction
Dynamic state of a car: 3+ foward veloctiy + raw rate
Kinematic state of an airplane, 6: x,y,z+ yaw rate+tilt+ roll
Dynamic state of an airplane:12

Monte Carlo Localization

particle filter
look at measurement, compute weight, sample, predict

office hour Q&A

Mahout: machine learning toolkit
sebastian: I never use toolkit. I’ve always coded straight in C++ or Java. More recently, as I aged, more in Matlab, which is really nice for rapid prototyping. It doesn’t give me the AI I want, but the tools to process the data, visualize data really quickly. So it’s fantastic to try out an idea.
next challenging?:
  • Object recognition.
  • Chess problem is only search problem.
  • perception problem. Do you understand the situation in fuzzy situations?
  • have an assistant. make everybody twice effective .
What hardware advancement expected to improve AI:
my 3-year old child is able to process huge amount of data. The way is: use existing computers and be really smart how to process massive amounts of data.
XX algorithms are hard to train?
Your laptop today was supercomputer of a a decade ago.
S: this is one of my frustration in scientific field, which in the past has looked into lots of small data sets, and very few have looked at very large data sets. As a result, the algorithms are mathematically beautiful and elegant, but they don’t scale.
I am pretty much self-taught. Look into the data first before applying any algorithm.
paths to professional in AI?
a lot of titles like computer scientist, game developer, research scientist, data scientist. No matter what the tools take, you’ve got to be able to get the job done.
The sexiest job in next 10 years will be statistician or machine learning person.