The number of particles and the max_timesteps for training
wenjunli-0 opened this issue · 0 comments
wenjunli-0 commented
Hi, I really like your work and I am having the following confuses. Could you please clarify them for me? Thanks very much.
- A SVPG particle is a sampling for the environment parameters, i.e. a sampling or a distribution of MainEnginePower if taking LLRandomized as an example. Then a single SVPG particle itself can have a distribution of MEP and generate a lot of corresponding environments. If this is the case, why would you set the number of particles for LLRandomized as N=10? Are you simply combining the 10 distributions of MEP together and then generate all of the environments? If this is the case, then can we set N=1, because one distribution MEP is the same as combining 10 distributions of MEP, and one distribution would be easier to learn.
BTW, I want to know whether you are arranging the particles as a curriculum, i.e. use the first particle to train the agent for the first 0-100k timesteps, and then use the second particles to train the agent for 100k-200k timesteps, and so on.
- Could you please tell me what's the max_timesteps you set in your experiment for LLRandomized? I hope to reproduce your results, but this parameter seems to be not mentioned in the paper or your codes.