Skip to content

Release v0.20
Compare
Choose a tag to compare
@takuseno takuseno released this 21 Dec 15:25
· 38 commits to master since this release
v0.20
80d9ab6

Support d3rlpy v0.41!

MINERVA now supports d3rlpy v0.41, which has a lot of progress from the v0.23 that is the previous dependency.

These are the some highlights related to MINERVA.

  • extremely fast mini-batch creation
  • extremely fast frame stacking for image observation
  • extremely fast N-step TD calculation
  • new metrics
  • etc

Algorithm Selection

MINERVA now provides many many algorithms for both discrete and continuous control datasets. You can choose an algorithm at the project creation dialog.

discrete algorithms

  • DQN
  • Double DQN
  • AWR
  • CQL
  • BCQ
  • SAC

continuous algorithms

  • DDPG
  • TD3
  • SAC
  • BCQ
  • BEAR
  • CQL
  • AWR
  • AWAC
  • PLAS

Of course, there is the Q function option to incorporate arbitrary algorithms with the powerful distributional Q functions.

Q functions

  • mean
  • Quantile Regression
  • Implicit Quantile Network
  • Fully parameterized Quantile Function
Assets 2
Loading