sashaperigo / cs224w-final Goto Github PK

Here's what I mean:

-for every primitive type:
for every primitive node of that type:
while primitive != TERMINAL_MARKER:
select the successor type
sample one successor node of that type according to some distribution (uniform if Erdos Renyi, distance-based if Watts-Strogatz)
record that successor node (or action)
primitive = successor
-return our actions

The samplers that we have currently take in percentages - we just need to either identify a percentage that returns a single node or change the percentage parameter to something else

the percentage you can use to get exactly one node back is:

float(1)/len(successor_indices_list)

Every time we build a Watts-Strogatz based graph (or do a Watts Strogatz random graph), we recompute the probability distribution from every primitive source to its successor list.

We should just cache these distributions - it'll save a ton of computation time for the random walks in particular.

I'm literally imagining a solution like:

--global--
CACHE = {}

def watts_strogatz_sampler():
if primitive_source not in CACHE:
recompute distribution & store it in CACHE
else: get distribution from CACHE
... use distribution

Maybe the one defined here: http://cyluun.github.io/blog/uninformed-search-algorithms-in-python

Except that we need to keep track of actions

Look at the shortest path code for an example of how to do things,

but overall, you should write some functions that run from main() and actually compute random walks and save their data to local CSVs