ML for Route Planning

[markjlascelles]

Hello,

I've been playing around with this project - being in Java makes it much easier to integrate with my existing project, thank you for writing it and making it available.

I've started off with cut down requirements, but am having trouble getting convergence towards the ideal solution.

To summarise

• take a target between 10m and 100m away from a vehicle
• allow a maximum acceleration/deceleration of +-1m/s-2
• require that the vehicle be travelling <1m/s when it reaches the target and that it arrives as quickly as possible

The ideal solution would I imagine be to accelerate as hard as allowed to about half way then decelerate hitting the target almost exactly at low speed.

I've modelled this with an NN that takes 2 parameters

• the current distance to the target
• the current velocity

I've mainly tried with Tanh activation functions - and the single output I've used as the new value for acceleration directly. Even with large generations (up to 10K) and many wide intermediate levels (up to 40x40), the output doesn't always get to the target, it's often going too fast/slow and oscillates around the final target (which adds time). Given I would like to add more requirements I was hoping I'd at least get close to ideal with the above.

I've tried all the activation functions provided, and also written new Relu and LeakyRelu AFs - and also allowed the output layer to have a different AF (Tanh) than the hidden layers (LeakyRelu) - but none seem to converge to the ideal.

Perhaps this NN is not suitable - perhaps a CNN is better or some of the other features mentioned in the issues section here might be needed.

Any thoughts or pointers on the above would be gratefully received.

Regards,
Mark