Might be worth exploring our own versions of ranked rewards. I don't expect that paper is a canonical standard at this point, so there might be further optimizations we could look into.
Do loss rewards need to be -1? Might be easier to make a loss 0 and win 1.
A couple ideas:
-
pareto front optimization
Here, we'd just check to see if a molecule is pareto-optimal, i.e., whether its dominated by any other recent solution currently in the game buffer.
select count(*) > 0 from game_table where buried_vol > %s and max_spin < %s and atom_type = %s
-
percentile scoring
Similar to ranked rewards but perhaps less sparse, just return 0->1 rewards depending on the percentile score of the game w.r.t a game buffer
-
anything else from the literature?
Might be worth exploring our own versions of ranked rewards. I don't expect that paper is a canonical standard at this point, so there might be further optimizations we could look into.
Do loss rewards need to be -1? Might be easier to make a loss 0 and win 1.
A couple ideas:
pareto front optimization
Here, we'd just check to see if a molecule is pareto-optimal, i.e., whether its dominated by any other recent solution currently in the game buffer.
select count(*) > 0 from game_table where buried_vol > %s and max_spin < %s and atom_type = %spercentile scoring
Similar to ranked rewards but perhaps less sparse, just return 0->1 rewards depending on the percentile score of the game w.r.t a game buffer
anything else from the literature?