Wiki
Clone wikiPyiB2c / Creating a behavior from scratch
Creating a behavior from scratch
The behavior module is the basic module inside an iB2c network. It inherits from the base class BehaviorTemplateClass and it can be created using the EmptyBehaviorClass. When the module is called:
- The activation variable is calculated.
- Function
self.calc_activity()is called. - Function
self.calc_transfer()is called. - Function
self.calc_rating()is called. - Variables
self.output,self.activity,self.ratingare returned.
Input and Output information
In order to make the network know what kind of variables the behavior is expecting, and to make the network architecture know its output, the variables "input_info" and "output_info" must be defined inside the module. Both variables are dictionaries with the following keys:
- Size: - np.array - sizes of each variables.
- Type: - np.array - short description of the variable type. To be freely defined by the user.
- Description: - list - description of each variable.
Example (see HeadToPointBehaviorClass):
#!python self.input_info = {'Size': np.array([np.zeros([1]), PoseV1Class(), np.zeros([2]), np.zeros([1])]), 'Type': np.array(['Time', 'Pose', 'Position2D', 'CoefAlpha']), 'Description': ['Simulation time [s]', 'Pose of the agent', '2D position of the target', 'Coefficient [0, 1] of velocity reduction']} self.output_info = {'Size': np.array([np.zeros([2])]), 'Type': 'Velocity2D', 'Description': '2D Velocity Command'}
This behavior is expecting 4 variables as inputs. The first input has dimension 1, its a time variable, and it is the simulation time in seconds. The second variable variable is a PoseV1Class class, which is a pose type: the pose of the agent.
Transfer function
Given an array of variables as inputs, this function calculates the output variables. It must be defined inside calc_transfer method.
Example (see HeadToPointBehaviorClass):
#!python def calc_transfer(self, inputs): # Get input variables t = inputs[0] pose = inputs[1] target_pos = inputs[2] alpha_v = inputs[3] # ------------------------------------------------------- # Calc output vector with the transfer function # ------------------------------------------------------- i_target = 0 r_a_t = [target_pos[i_target][0] - pose.X[0], target_pos[i_target][1] - pose.X[1]] psi_a_t = np.arctan2(r_a_t[0], r_a_t[1]) delta_psi = psi_a_t - pose.psi # Correction [-pi,+pi] self._delta_psi = np.arctan2(np.sin(delta_psi), np.cos(delta_psi)) # Calculation of the angular velocity w = self.P*self._delta_psi/np.pi delta_t = t - self._t self._t = t psi_c = pose.psi + delta_t * w # ------------------------------------------------------- v_c = np.zeros(2) v_c[0] = np.sin(psi_c) v_c[1] = np.cos(psi_c) v_c = alpha_v*self.v_n * v_c self.output[0] = v_c
Activity and rating function
The activity and the rating variables are calculated.
Example (see HeadToPointBehaviorClass):
#!python def calc_activity(self): self.activity = self.activation def calc_rating(self): self.rating = 1 - np.abs(self._delta_psi) / np.pi
Final step
Save the module inside the Behaviors folder and add the information to the behaviors_dictionary.xml file, for instance:
#!xml <?xml version="1.0"?> <data> <behavior name="HeadToPointBehavior"> <class_name>HeadToPointBehaviorClass</class_name> <module_name>HeadToPointBehaviorClasses</module_name> <activated>1</activated> </behavior> </data>
Updated