SMLG (Statistical Machine Learning Group) Discussion Forum

Some rough pseudocode of what I am implementing:

Code: Select all

MCMC-ORDER-SEARCH(dataset d, max_parents k, max_time t):
[1. INITIALIZATION:]
  od_0 = random_order( d )
  s_0 = log_score( od_0, d, k )
  results = { (od_0, s_0) }
  end_time = system_time() + t

  while ( system_time() < end_time )
[2. PROPOSAL:]
    od*_{i+1} = random_swap( od_i )
    s*_{i+1} = log_score( od*_{i+1}, d, k )

[3. ACCEPT-REJECT:]
    if ( rand[0,1] < exp( s*_{i+1} - s_i ) )
      od_{i+1} = od*_{i+1}
      s_{i+1} = s*_{i+1}
    else
      od_{i+1} = od_i
      s_{i+1} = s_i
    end

    results = { results, (od*_{i+1}, s*_{i+1}) }
  end


The main detail I am not too sure about is the last step. In a sampling algorithm, we would store od_{i+1} since that is the step we made. In this version, it seems that we would ignore duplicates (i.e. it wouldn't be useful to attempt to add od_{i+1} when od_{i+1}=od_i).

Michael Cho wrote:Some rough pseudocode of what I am implementing:

Code: Select all

MCMC-ORDER-SEARCH(dataset d, max_parents k, max_time t):
[1. INITIALIZATION:]
  od_0 = random_order( d )
  s_0 = log_score( od_0, d, k )
  results = { (od_0, s_0) }
  end_time = system_time() + t

  while ( system_time() < end_time )
[2. PROPOSAL:]
    od*_{i+1} = random_swap( od_i )
    s*_{i+1} = log_score( od*_{i+1}, d, k )

[3. ACCEPT-REJECT:]
    if ( rand[0,1] < exp( s*_{i+1} - s_i ) )
      od_{i+1} = od*_{i+1}
      s_{i+1} = s*_{i+1}
    else
      od_{i+1} = od_i
      s_{i+1} = s_i
    end

    results = { results, (od*_{i+1}, s*_{i+1}) }
  end


The main detail I am not too sure about is the last step. In a sampling algorithm, we would store od_{i+1} since that is the step we made. In this version, it seems that we would ignore duplicates (i.e. it wouldn't be useful to attempt to add od_{i+1} when od_{i+1}=od_i).

The prior idea involves in random_order() and random_swap(). User can specify their pror knowledge on the ordering search. For example, the user can specify any pairwise relationships using the following syntax for each line:

node1, node2, prior value

which represents that node1 should proceed node2 with prior probability of the given prior value. For example, if a user specifies the following lines (we ask users to specify probability that is always higher than 0.5, i.e., you can specify the first line as gene1, M_gene1, 0.0 as well):

M_gene1, gene1, 1.0
gene2, gene3, 0.7

Then whenever you generate a random_order() or random_swap(), you will make sure M_gene1 does not end up in a lower order than gene1. Also, when generating a random_order() or random_swap(), if gene2 is in higher order of gene3 then proceed, however, if gene2 become lower order of gene3, then only accept it when you toss a bias coin with getting a head of (1.0-0.7)/0.5 = 0.6 and when the head comes out. Please double check the math.

Michael, is this code anywhere in the Git repository?

Built on the pseodocode provided

vlee14 wrote:Built on the pseodocode provided

It seems Michael implemented the following module:

log_score( od, d, k )

Where od is an order, d is a dataset, and k is the maximum number of parents. Look for that module in Git library.

Worked on the order search code as well as the corresponding header file.

Still working on the random swap module.

vlee14 wrote:Still working on the random swap module.

What part of the code is taking this long? Please post it here and explain why it requires more than eight hours to work on it. You need to be more efficient in implementing modules.

Finished generating a random order and implementing the random swap method and then scoring both orders.

vlee14 wrote:Finished generating a random order and implementing the random swap method and then scoring both orders.

What are left to complete the MCMC order search?