Fix to how we select the normalize_state HPO parameter. We always want to normalize the representations to [-1, 1] (since the metrics can be arbitrarily large) for neural network stability (#21).
Added an error check such that when we clip the gradient, it'll throw if it's already bad for clearer error checking (#21).
For a bit more "backwards" compatibility, added back the structure_normalize option where we use the latent projection directly and not rely on the observation based normalization.
When a query times out, there might be additional DDL afterwards (TPC-H Q15 is [create view, select, drop view]). We need to actually execute the drop view cleanup if executing the select exceeds the workload timeout (#23).
Small fixes for #21 and #23.
normalize_stateHPO parameter. We always want to normalize the representations to[-1, 1](since the metrics can be arbitrarily large) for neural network stability (#21).structure_normalizeoption where we use the latent projection directly and not rely on the observation based normalization.[create view, select, drop view]). We need to actually execute thedrop viewcleanup if executing the select exceeds the workload timeout (#23).