|
Scaling up of
lexicon and action repertoire in Cognitive Agents
Description:
Classification and categorization of actions / building
sensorimotor concept-models
Let’s first
consider having 112 different actions, some inspired by an
alphabet system (the semaphore flag signalling system) see
figure 2. We have collected data on the posture of robots using
6 features. The object input data consist of the 6 angles of
each, left arm and right arm joints. (Shoulder, upper arm and
elbow) The agents first have to learn to classify these actions;
at this stage we are using a multi-dimensional MFT algorithm
with 112 fields randomly initialized. Figure 1 shows that the
model is able to correctly identify the different actions.
Although the simulation initially dealt with 112 actions the MFT
algorithm was able to categorize to approximately 95% successful
matching. Therefore there was a slight reduction in the number
of completed actions. Figure 3
shows our system consisting of two simulated agents (teacher and
learner) embedded within a virtual simulated environment (Using
Open Dynamic Engine).
Results:
Figure 1:
Time evolution of the fields with 6 features being used as
input: 112 different actions.
Figure 2:
Few examples of type of behaviour used for the classification
and categorization of actions. (here the semaphore alphabet)
Figure 3:
Teacher and learner before action is learnt and after
Up -
Back
|
