change map after few successful rounds of map for model generalization

[kaustubholpadkar]

In the existing code, new generations are not created once the particle successfully learns the current map and those particles in the map completing it triumphantly! indefinitely. In this case, we think that the particle has learned to navigate in the maps, but it is generally overfitted to the current map only. So, to make the particle generalize enough that it can navigate in a large variety of maps, I have added a logic to build a new map once the particle has successfully navigated the current map and made the round-trip 5-6 times. Thus, periodically the model will explore many maps and try to adjust weights of neural network in a way that helps it navigate the variety of maps.

[shiffman]

Thank you for this!

[jonnytest1]

@kaustubholpadkar 5 rounds should only give 300 fitness (5 rounds times 60 checkpoints) with the current setup oO , maybe extract total from track generation and make it 5*total ?

https://github.com/CodingTrain/NeuroEvolution-Vehicles/pull/6

[kaustubholpadkar]

Thank you for this!

My pleasure!

[kaustubholpadkar]

@kaustubholpadkar 5 rounds should only give 300 fitness (5 rounds times 60 checkpoints) with the current setup oO , maybe extract total from track generation and make it 5*total ?

6

Thats true. We can make it more dynamic that way.