Code

Add Lookahead+RAdam optimizer#416

Code

Merged

kengz merged 19 commits intomasterkengz/SLM-Lab:masterfrom lookaheadkengz/SLM-Lab:lookaheadCopy head branch name to clipboardSep 23, 2019

+178-89

Commits

Commits on Sep 5, 2019

• implement lookahead optimizer

  kengz

  committed
  42742c3

  Copy full SHA for 42742c3

• try LookAhead on ppo roboschool

  kengz

  committed
  fc08261

  Copy full SHA for fc08261

• use lookahead in a2c

  kengz

  committed
  a8a0a4a

  Copy full SHA for a8a0a4a

Commits on Sep 6, 2019

• replace all roboschool benchmark to use Lookahead

  kengz

  committed
  9d76af7

  Copy full SHA for 9d76af7

• remove invalid swingup env

  kengz

  committed
  54de6b9

  Copy full SHA for 54de6b9

Commits on Sep 7, 2019

• make RAdam and Lookahead global

  kengz

  committed
  6e9a8fd

  Copy full SHA for 6e9a8fd

• share optimizer based on method availability

  kengz

  committed
  a417124

  Copy full SHA for a417124

• set properties in RAdam and Lookahead for a3c share_memory

  kengz

  committed
  3f61399

  Copy full SHA for 3f61399

• update RAdam from author, fix adap_lr as scalar

  kengz

  committed
  a45373a

  Copy full SHA for a45373a

• use defaults to track lookahead

  kengz

  committed
  add6a17

  Copy full SHA for add6a17

• commit working a3c gae cartpole spec

  kengz

  committed
  37570d9

  Copy full SHA for 37570d9

• handle inner optimizer setting for global nets

  kengz

  committed
  8fb302b

  Copy full SHA for 8fb302b

Commits on Sep 13, 2019

• lower lr for sac lookahead roboschool

  kengz

  committed
  108c24b

  Copy full SHA for 108c24b

• Merge remote-tracking branch 'origin/master' into lookahead

  kengz

  committed
  d64f24e

  Copy full SHA for d64f24e

Commits on Sep 16, 2019

• revert SAC specs back to old

  kengz

  committed
  0f1b5b3

  Copy full SHA for 0f1b5b3

Commits on Sep 22, 2019

• tune sac harder

  kengz

  committed
  d1d9e59

  Copy full SHA for d1d9e59

Commits on Sep 23, 2019

• Revert "tune sac harder"

  Show description for 63bb516
  kengz

  committed
  63bb516

  Copy full SHA for 63bb516

• update benchmark tables

  kengz

  committed
  e5988f0

  Copy full SHA for e5988f0

• add legend

  kengz

  committed
  3be6fd6

  Copy full SHA for 3be6fd6