Noise Benchmarking, LSTM sb3_zoo hyperpar, Issue #1 #9
Conversation
…nor changes to envs
…py now accepts new yaml syntax & old (TBD: re-standardize)
…sheries into benchmarking
|
Converted to draft to solve Issue #1 |
felimomo
changed the title
|
Updated based on our meeting! Also, side note: now |
There was a problem hiding this comment.
@felimomo fantastic work here! A few comments in line. Basically I think let's just add the harvest_vul in now as well, since it looks odd to have something explicitly called survey_vul being used inside the harvest function, and then merge this.
I think we're due for a refactor of the Asm code, so that it's easier to read and easier to customize. (most importantly, asm_2o should re-use all the code except for an override observation method, and similarly for escapement with an override step method I think). but let's tackle that in a separate PR.
| @@ -174,7 +173,7 @@ def initialize_population(self): | |||
|
|
|||
| # leading array calculations to get vul-at-age, wt-at-age, etc. | |||
| for a in range(0, p["n_age"], 1): | |||
| vul[a] = 1 / (1 + np.exp(-p["asl"] * (p["ages"][a] - p["ahv"]))) | |||
| survey_vul[a] = 1 / (1 + np.exp(-p["asl"] * (p["ages"][a] - p["ahv"]))) | |||
There was a problem hiding this comment.
once we have a harvest_vul, it's going to need to be initialized here too. Presumably it's identical but has different values for p["asl"] and p["ahv"]. (We'll need @CarlJwalters or @ChrisFishCahill to figure out some reasonable choices there)
src/rl4fisheries/envs/asm.py
Outdated
| self.abar = sum(p["vul"] * np.array(p["ages"]) * n) / sum(n) | ||
| self.wbar = sum(p["vul"] * n * p["wt"]) / sum(n * p["wt"]) | ||
| self.abar = sum(p["survey_vul"] * np.array(p["ages"]) * n) / sum(n) | ||
| self.wbar = sum(p["survey_vul"] * n * p["wt"]) / sum(n * p["wt"]) |
There was a problem hiding this comment.
same as above, eventually this should be harvest_vul
In this branch:
Asm().sigma) with a script input, added a bash script to train agents for several values of noise simultaneously.Next steps: sb3_train_v2 is more general than sb3_train, however its syntax is a bit more involved and might need refactoring in order to reach sb3_train's level of clarity. I'll leave this for the next pull request -- the code is fully functional right now, although in a 'quick-n-dirty' state in some parts.