Define_Graph: Graph for Network Loss according to A3C.

In MartinKies/USLR: Reinforcement Learning with R

Description

Defines a tensorflow graph specifiying the loss calculation according to the A3C algorithm.

Usage

Define_Graph(model, model.par, game.object)

Arguments

model	A Neural Network e.g. as given by Setup.Neural.Network.A3c or Setup.Neural.Network.A3c.LSTM.
model.par	A list with parameters to set up the Network e.g. as given by Get.Def.Par.Neural.Network.A3C or Get.Def.Par.Neural.Network.A3C.LSTM.
game.object	A Game Object (list) as defined by Get.Game.Object.<NAME>.

Details

Returns a list of tensors with the following items:

• gradient.clipped Calculated gradients of the network weights.

• loss.policy Calculated loss of the policy.

• loss.value Calculated value loss.

• loss Calculated total loss of the network.

• s_ Placeholder for states.

• a_ Placeholder for actions.

• r_ Placeholder for rewards.

• adv_ Placeholder for calculated advantages.

MartinKies/USLR documentation built on Nov. 10, 2019, 5:24 a.m.