MultiModalPolicyOptions<T>
Options & Configuration · AiDotNet.ReinforcementLearning.Policies
Configuration options for multi-modal mixture-of-Gaussians policies: MultiModalPolicy’s
#ctor constructor builds its network from them.
Inheritance: ModelOptions → MultiModalPolicyOptions<T>
For Beginners
Use this when there can be more than one right answer (go around an obstacle on the left or on the right). A single Gaussian would average the two and drive into the obstacle.
How It Works
A mixture density network (Bishop, 1994) head: for each of NumComponents components the network
outputs a mixing logit, a mean per action dimension and a log standard deviation per dimension, so a state can
have several distinct good actions instead of one Gaussian averaging between them.
Constructors
| Constructor | Summary |
|---|---|
MultiModalPolicyOptions() |
Properties
| Property | Summary |
|---|---|
IExplorationStrategy<T> ExplorationStrategy { get; set; } |
Gets or sets the exploration applied to sampled actions during training. |
ILossFunction<T> LossFunction { get; set; } |
Gets or sets the loss the policy network is trained with when it is fitted directly. |
int ActionSize { get; set; } |
Gets or sets the number of action dimensions. |
int NumComponents { get; set; } |
Gets or sets the number of Gaussian components in the mixture. |
int StateSize { get; set; } |
Gets or sets the number of features in an observation (the network’s input width). |
int[] HiddenLayers { get; set; } |
Gets or sets the width of each hidden layer of the policy network, in order. |