Hi, thanks for this work.
I wonder how are the clauses in the threefish env (nsfr/nsfr/data/lang/bigfish/bigfish_bsrf1/clauses.txt) generated?
I tried to use the provided ppo model (src/models/bigfish/ppo/ppo.pth) and the file collect_data.py to generate the data.json.
I used scoring to beam_search the clauses, while the search ones are different from your provided clauses and are very reasonable, e.g.,:
down_to_eat(X):-high_level(02,01),low_level(01,02),type(01,agent),type(02,fish).
How can I get the correct clauses?
Hi, the error you obtain might be due to the fact that you tried a refinement of 1 directly (which is very brittle).
Have you tried with higher selection numbers?
Hi, thanks for this work. I wonder how are the clauses in the threefish env (nsfr/nsfr/data/lang/bigfish/bigfish_bsrf1/clauses.txt) generated? I tried to use the provided ppo model (src/models/bigfish/ppo/ppo.pth) and the file collect_data.py to generate the data.json. I used scoring to beam_search the clauses, while the search ones are different from your provided clauses and are very reasonable, e.g.,:
down_to_eat(X):-high_level(02,01),low_level(01,02),type(01,agent),type(02,fish).
How can I get the correct clauses?Thanks.