Skip to content

A virtual whiteboard so I don't forget the ideas that come to me

License

Notifications You must be signed in to change notification settings

dscolby/Whiteboard

Folders and files

NameName
Last commit message
Last commit date

Latest commit

 

History

14 Commits
 
 
 
 
 
 

Repository files navigation

Whiteboard

This is a virtual whiteboard for me to flesh out some ideas I've had that may or may not be work well and make good starting points for research. The idea is that eventually I will get around to putting them to the test or realizing that they won't actually work. If you've got a similar idea, think of some challenges I might run into or something that I'm overlooking, or want to work together, feel free to open an issue, submit a pull request, or get in touch.

Categories

Causal Discovery

Reinforcement Learning

Game Theory and Other Sequential Decision-making Formalizations

Causal Discovery

Causal Discovery as a Bandit Problem

  1. Problem: CI test/score function scales with datapoints
    1. Solution: sample random subsets of data and aggregate over samples
      1. LLN implies that average will converge to true value for whole dataset
  2. Problem: too many DAGs to enumerate
    1. Solution: causal discovery as an (infinitely-armed) bandit problem
      1. Each DAG is an arm
      2. Use score function/CI tests on subsets of data as score for arm (DAG)
        1. If using CI tests, need to apply Meek's rules afterwards.
      3. One potential algorithm is GP with minimal regret acquisition function
        1. Could use NN version of GP or infinitely-armed bandit algorithm, various algorithms already exist
        2. Problem: need to convert DAGs to unique arms, ideally without constructing each DAG
          1. Solution: bijection between DAGs and integers
            1. Adjacency matrix of DAG is lower triangular 1s and 0s

            2. If entry is not 1, it is 0 so really only need to focus on one or the other

            3. Concatenate element positon of 1s (row number, colum number) to get unique integer

            4. Example:

                           |    *     | Column 1 | Column 2 |  Column 3 |
                           |  :---:   |   :---:  |   :---:  |   :---:   |
                     A =   |   Row 1  |     1    |     0    |     0     |
                           |   Row 2  |     0    |     1    |     0     |
                           |   Row 3  |     0    |     0    |     0     |
              
                     f(A) = 1122
              
            5. Can enumerate over these integers without explicitly constructing DAGs or adjacency matrices

  3. With arm IDs and way to calculate (stochastic) rewards each time we pull an arm (DAG), we can use bandit algorithm to find the best arm!
  4. Can also think of this as adaptive MCMC where probability of moving to next state (arm) is UCB of one arm/sum of UCB for all arms.

Causal Discovery via Tabu Search

  1. We can use the same approach as above to calculating scores or independence tests
  2. Then run Tabu Search over the vectors
    1. Tabu search searches a neighborhood and looks for the best solution
    2. Then searches the neighborhood of the best solution and repeats
    3. Seen vertices are added to a tabu list to not search
      1. They get removed when the tabu list becomes full

Causal Discovery via MCMC

  1. Same approach as above to calculating scores or independence tests
  2. Same approach for assigning unique numbers to DAGs
  3. Use Metropolis-Hastings MCMC to sample over graphs
    1. Ideally want to make it converge faster
      1. Maybe use temperature schedule
      2. Could also use some kind of exploration bonus
        1. Want to encourage exploration early on but quicly converge to the best DAG(s)
          1. Maybe use active learning or similar approach to Ready Policy One paper

Reinforcement Learning

Global Active Policy Search

  1. Core Problems/Setup
    1. In theory RL can be really simple, just need to find parameters to optimize a policy
    2. Linear policies seem to work well for a lot of tasks, like in the ARS paper
    3. Using policy gradients is complicated and sample inefficient
  2. Solution: Use global search/BO to find best parmeters
    1. Already done by Mueller et. al, but uses warm starts
    2. Use linear policy parameters as inputs and predict cumulative expected rewards with UCB
    3. Use L2 norm between actions from initial policy states to compare policies and add them to Materne kernel
      1. Doesn't work well without warm start policy
    4. How to deal with kernel issue
      1. Use learned kernel
        1. Use DKL or similar approximator to learn a kernel
          1. Also scales better than regular BO
        2. Could also use non Bayesian verison
          1. Predict mean and use dropout to get uncertainty estimates/UCB
          2. GO-UCB

Game Theory and Other Sequential Decision-making Formalizations