grf-labs/policytree

Fix test case for approximate tree search

Closed this issue · 6 comments

" tree search with approximate splitting works as expected ": fix this test case as it fails sometimes.

Can you say which test fails?

The test as cited above ("tree search with approximate splitting works as expected")..

That test has 8 checks, are you saying all 8 checks always fail?

oh; within that.

expect_true(reward.skip2 > 1.5 * reward.colmax)

This fails probabilistically. It passed with my current set of changes which are only aesthetically different from the last. Maybe we could hardcode the seed so that it does not fail probabilistically (at least on machines with Linux?).

Does this show up as 8 tests or 1 test in the testing console (for my information)?

It should show up as failed: "tree search with approximate splitting works as expected" then show which checks within that failed.

How about just changing it to expect_true(reward.skip2 > reward.colmax) which is just a loose check that it is not doing anything terrible (the behavior tested should be robust enough that it holds regardless of seed)