tree b6285683daac0973aa4053a59164d54eb58765a9
parent 40373da34f9a3b4bd7103c4e8d442ecf5fdafd3e
author Austin Schuh <austin.linux@gmail.com> 1729982358 -0700
committer Austin Schuh <austin.linux@gmail.com> 1729989132 -0700
gpgsig -----BEGIN PGP SIGNATURE-----
 
 iQGzBAABCgAdFiEExcNA7Sy+j6Ly5hA4QVYw08sOnWYFAmcdihsACgkQQVYw08sO
 nWYDYgv9E5v6S76sMNL0hgse9uIgX9qZ346lbg0UueQGdXeGOsB9xmu7Ni/PPWl8
 E11Mn0HJItsDnX6CipHZHeSVT75mRZHaHPgVsAyw7dMfACrCujkmC7z6vpZ5Zwl1
 HpeTZcrVfLamflt8AisqlPRvBKHnBevV2F+76JN0saLUfGyFS0S0+jldtKbW6Ypj
 9B3gCB2A8GT1J4h3UpPji8+D9yuutaY6aMqmkGB+KYiScjSh9fLzyZd4OOGqmdoB
 C2/Z3xWLvrAjGOEOsWCJxB+mvk/PzO4PBylAevrs0Wd/VyNNZCHONN6/jdG0A/2Z
 rHV8jXUCMc/c4XzzdXbkEgyskGIyNwffx38dLiMKKW7imL6+d+UO5CEG4xGau2zV
 Ks6oPtxESfnfXmc0KzA0przAV9lVQ3tryf+Y2Y753lWRMpgOltvgeOBUF6aJewRe
 T/2KHBfOPxAosMhHXv74OAQDTgYfWNesxP4jIDQ8cIJC5cTYU+a1v1npdq0CPp1F
 vExtVD5z
 =eDjt
 -----END PGP SIGNATURE-----

Switch to using tensorflow_probability

This is technically less accurate, but is a lot less to validate since
we are having trouble getting this to converge.

To try to adhere to the "unit gaussian" assumption that neural networks
like, make the physics account for the action limit instead of the actor
network.  We can undo this after things work if we find it doesn't make
it converge better.

Also, switch from talking about "cost" to "reward" to adhere to other
implementations closer too.

Change-Id: I1b8df0a72946020b3e3da5cf46b3787e38cb94c9
Signed-off-by: Austin Schuh <austin.linux@gmail.com>