Can we use other environments (such as Treechop) to train the agent before training on the Diamond env?
Absolutely! During training, all MineRL
environments are available (and steps in those environments are counted in the 8 million step limit just like steps in MineRLObtainDiamond-v0
)
Just remember for evaluation, your agent will only be scored in MineRLObtainDiamond-v0
and other environments will not be available
@BrandonHoughton In other words, during the training stage, we can use >>> 8M
step in other’s environments, besides the MineRLObtainDiamond-v0
.
For example, we can use 10 million step in Treechop
, and we can use 20 million step on MineRLObtainIronPickaxe-v0
,… etc, and we just make sure that step in ObtainDiamond-v0 < 8 million
?
Is it correct? Thanks
@weel2019 I believe this is not the case. The total number of steps, no matter what environment you use to train your agent, should not exceed 8M steps, if I understood Brandon’s comment correctly.
Thanks, @jazzpiano3 From the rule,it’s 8M total (for all the environments). I thinks i would have misunderstood about above info.
I checked the starter_kit. The parser
always checks if the environment is “ObtainDiamond”. Are we allowed to modify that part in order to train on other envs?
All environments are available for training! For testing the parser will fix the environment to MineRLObtainDiamond-v0