Question L4-C1

The question is about setting the weights of a simple neural network with two neurons to do obstacle avoidance of a differential drive robot, basing the control on the five front proximity sensors and controlling the wheels speeds with the output of the neurons. On the right are illustrated the structure of the network with some weights (A) and the configuration of symmetrical obstacles (B). The question is: "To implement a local obstacle avoidance, one can use a neural network with two simplistic neurons (no output function) and use it to connect sensors to motor speed targets, as illustrated in the schematics A on the right. If the weights of the network are symmetrical, the robot faced to a symmetrical obstacles (picture B.) will be
unable to avoid. One of your colleagues proposes to have a non-symmetrical weight to avoid this problem. What sentences are right (several possible answers)?". Three possibilities are proposed:

Answer A: "The asymmetrical weights, if they implement avoidance on both directions (on the right and on the left) will not solve the problem of local minima."
This answer is correct: While asymmetrical weights provide partial directionality, if the avoidance can be done in both directions there will be always a point of symmetry between the two sides where the robot will be unable to avoid.
In the explanation of the student we would like to see that they understand that this solution only moves the problem to another point of equilibrium but is not solving the problem.

Answer B: "The asymmetrical weights, if set correctly, will solve the problem."
This answer is correct: If the weights will prioritise only one direction among the two, there will be no point of equilibrium anymore and therefore the robot will be able to avid the obstacle. 
In the explanation of the student we would like to see that they understand that the option here is to choose only one direction of avoidance.

Answer C: "Changing the weights is like changing the basic vectors of a potential field approach, the reasoning can be made using this tool also."
This answer is also correct: In the potential field method, obstacles generate repulsive forces that influence the robot’s movement, guiding it away from the obstacle by weighting 2-dimensional vectors. By adjusting the weights of the two neurons, you're essentially doing the same work than weighting the vectors. The method and result is the same, altering the strength and direction of these "forces," (in the potential field method) and controlling how the robot reacts to obstacles. The number of parameters is identical, only the interpretation changes.
In the explanation of the student we would like to see that they understand that the two approaches are just different interpretations of the same weighting mechanism of the sensors to generate motor speeds.