SmartEngine
1.6.0
|
Data used to construct an ICuriosityModule instance More...
#include <CuriosityModule.h>
Public Attributes | |
IContext * | context = nullptr |
Context the constructed graph will belong to. Only graph nodes of the same context can be connected. More... | |
IGraph * | graph = nullptr |
The agent graph for the constructed curiosity graph. More... | |
const char ** | actionNodeNames = nullptr |
The list of action output nodes in the agent graph that we will track. More... | |
int | actionNodeNamesCount = 0 |
The size of the action node name array. More... | |
int | observationFeatureDimension = 16 |
Internally, the observation input is turned into a feature space. This value specifies the dimension of that feature space. More... | |
float | rewardMultiplier = 0.01f |
The factor that is multiplied with the generated rewards. More... | |
float | maxIndividualReward = 0.0f |
The maximum allowed reward for a single observation. More... | |
bool | normalizeRewards = false |
True if we should normalize the rewards. This will turn large and small rewards into a normal curve. More... | |
float | minThresholdStdDev = 0.0f |
Reward values less than this many standard deviations away from a value of 0.0 will be set to 0.0. More... | |
float | inverseForwardLossWeight = 0.2f |
A sliding [0..1] value that specifies how much we weight training the internal forward network (Predicting the next state from the current state and actions) versus the internal inverse network (predicting the action from the current state and next state). More... | |
int | sequenceLength = 1 |
The sequence length to use when using an agent graph that requires stepping (such as the inclusion of an LSTM). More... | |
GradientDescentTrainingInfo | trainingInfo |
The parameters used to train the internal curiosity graph. More... | |
Public Attributes inherited from SmartEngine::ResourceCInfo | |
const char * | resourceName = nullptr |
Optional resource name that will be used with Load() and Save() if no other name is provided. More... | |
Data used to construct an ICuriosityModule instance
const char** SmartEngine::CuriosityModuleCInfo::actionNodeNames = nullptr |
The list of action output nodes in the agent graph that we will track.
int SmartEngine::CuriosityModuleCInfo::actionNodeNamesCount = 0 |
The size of the action node name array.
IContext* SmartEngine::CuriosityModuleCInfo::context = nullptr |
Context the constructed graph will belong to. Only graph nodes of the same context can be connected.
IGraph* SmartEngine::CuriosityModuleCInfo::graph = nullptr |
The agent graph for the constructed curiosity graph.
float SmartEngine::CuriosityModuleCInfo::inverseForwardLossWeight = 0.2f |
A sliding [0..1] value that specifies how much we weight training the internal forward network (Predicting the next state from the current state and actions) versus the internal inverse network (predicting the action from the current state and next state).
float SmartEngine::CuriosityModuleCInfo::maxIndividualReward = 0.0f |
The maximum allowed reward for a single observation.
Specify 0 to not cap a maximum value
float SmartEngine::CuriosityModuleCInfo::minThresholdStdDev = 0.0f |
Reward values less than this many standard deviations away from a value of 0.0 will be set to 0.0.
Specify 0 to not set a minimum standard deviation.
bool SmartEngine::CuriosityModuleCInfo::normalizeRewards = false |
True if we should normalize the rewards. This will turn large and small rewards into a normal curve.
int SmartEngine::CuriosityModuleCInfo::observationFeatureDimension = 16 |
Internally, the observation input is turned into a feature space. This value specifies the dimension of that feature space.
float SmartEngine::CuriosityModuleCInfo::rewardMultiplier = 0.01f |
The factor that is multiplied with the generated rewards.
int SmartEngine::CuriosityModuleCInfo::sequenceLength = 1 |
The sequence length to use when using an agent graph that requires stepping (such as the inclusion of an LSTM).
GradientDescentTrainingInfo SmartEngine::CuriosityModuleCInfo::trainingInfo |
The parameters used to train the internal curiosity graph.