The settings of the DQN-Agent
Properties:
| Name | Type | Attributes | Description |
|---|---|---|---|
learningRate |
number
|
The learning rate |
|
discountFactor |
number
|
The discount factor |
|
nnLayer |
Array.<number>
|
The size of the neurlal network layer |
|
replayMemorySize |
number
|
The replay memory size |
|
batchSize |
number
|
The batch size |
|
replayMemoryInitSize |
number
|
The initial needed size of the replay memory |
|
epsilonStart |
number
|
The epsilon start value |
|
epsilonEnd |
number
|
The epsilon end value |
|
epsilonDecaySteps |
number
|
The number of epsilon decay steps |
|
activateDoubleDQN |
boolean
|
<nullable> |
Use a double DQN setup for learning (recommended) |
updateTargetEvery |
number
|
<nullable> |
After how many steps to synchronize the target network with the local network when the double DQN is active |
hiddenLayerActivation |
string
|
<nullable> |
The hidden layer activation function to use. (recommended: 'relu'). See the tensorflow.js documentation for available activation functions. |
layerNorm |
boolean
|
<nullable> |
Whether to use layer normalization. This slows down training but may improve training results. |
kernelInitializerSeed |
number
|
<nullable> |
The seed to use for kernel initialization. This improved the reproducability of training results. |