Lecture notes taking during CS238, decision making. Stanford Intelligence Systems Laboratory (SISL: planning and validation of intelligent systems).

Big Ideas

Themes

1. There’s a principled mathematical framework for defining rational behavior
2. There are computational techniques that could lead to better, and perhaps counter-intuitive decisions
3. Successful application depends on your choice of representation and approximation
   • you typically can’t solve mathematical models exactly
   • so, we have to rely on good models of approximations
4. The same computational approaches can be applied to different application domains
   • the same set of abstractions can be carried through life
   • send Mykel a note about how these topics about where this stuff is applied

These algorithms drive high quality decisions on a tight timeline. You can’t fuck up: people die.

A decision network is a Baysian Network which is used to make decisions based on optimizing utility.

To solve a problem, we iterate through all possible decision parameters to find the one that maximizes utility.

Nodes

1. chance nodes: random variables — some inputs we can observe, some are latent variables we can’t observe — circles
2. action nodes: what we have control over — squares
3. utility nodes: output, what the results would be; we typically sum utilities together if you have multiple of them — diamonds

Edges

1. conditional edge - arrows to chance nodes: conditional probability edges
2. informational edge - arrows to action nodes: this information is used to inform choice of action
3. functional edge - arrows to utility nodes: computes how the action affects the world

Example

For \(U\), for instance, you can have a factor that loks ilke:

a student approach to learning where learning outcomes are driven by student’s own experience to deeply drive educational results independenlty

deep learning is MLE performed with neural networks. A neural network is many logistic regression pieces (sic.?) stack on top of each other.

We begin motivating this with trying to solve MNIST with logistic regression. What a time to be alive. After each layer of deep learning, we are going to use a layer of “hidden variable”, made of singular logistic regressions,

Notation:

\(x\) is the input, \(h\) is the hidden layers, and \(\hat{y}\) is the prediction.

Facts

1. Everybody writes bugs
2. Debugging sucks

Defensive Programming Tools + Techniques

• Use language features
• Specs, documentations, Test-Driven Development, unit testing
• Fail fast and loudly
• Systematic debugging
• Investing in tools

Use Language Features

• Descriptors: static, final, pub./priv.
• Type checking: prevent type errors
• Automatic array bounds checking
• Memory management
• Compiler optimization

Key idea: know what language features are available, why/when to use them. don’t work against the language in circumventing them

Specs, Docs., TDD, Unit Tests

• How should it work: specs
• How does it work: docs
• How will I know it works: TDD
• How do I know it still works: unit tests

These all force you to think about your code before!! you write it so then you can correct them as soon as possible.