1 minute read

Observing the large training time that the algorithm needs to learn to solve the problem when the iteration speed is limited (since the lower the speed, the fewer cases it observes per second and, therefore, the less is the training), an alternative that could be faster has been chosen. In this case it has been decided not to limit the iteration speed of the algorithm and to try different values for the number of bins and maximum and minimum values of the rotation and linear velocities. First, this approach was tested during inference while modifying the Q table on the same scale to see if this could improve the car times, but during this week it has been impossible to calibrate it or at least, no results have been observed. Then, we proceeded to run a new training modifying only the maximum and minimum values of the speeds, however, the car is not able to stabilize correctly and there are certain scenarios that it is not able to solve. In the end, an approach has been reached where a number of bins is determined for each odd speed, so that there is always the value of 0 turn and thus achieve to center the car in the straight lines, and leave the same number of bins for the observations, since these have been considered to be sufficient to solve the problem with previous parameterizations. Currently, the algorithm is training with a number of bins of 7 for the linear velocity, 7 for the angular velocity and 10 for the observations. As maximum values of linear and angular velocity both have been set to 10 and as minimum values, the first one to 0 and the second one to -10.