April 2018
Intermediate to advanced
334 pages
10h 18m
English
The first part of the state of the agent is defined by the visual features extracted by using two models, which are:
These two variations are explained in the Model and Training section that will follow.
The second part of the state of the agent is the memory vector, which captures the actions of the past four time steps the agent took in order to search for the object. At each time step, there are six possible actions (described in the section to follow). Therefore, the memory vector has 4*6 = 24 dimensions. This memory vector has been found useful to stabilize the search trajectories.