search

aavek / Aeolus-Ocean

An all-weather, day-and-night, collision avoidance simulator that can be implemented as a digital twin for the autonomous COLREG-compliant navigation of maritime vessels.
BSD 3-Clause "New" or "Revised" License
30 stars 2 forks source link

How to deploy other reinforcement learning path planning algorithms #2

Open ChirsDejong opened 1 year ago

ChirsDejong commented 1 year ago

I would like to express my gratitude for providing this remarkable platform. I am highly interested in How to deploy other reinforcement learning path planning algorithms and I'm eager to explore its capabilities on the current platform. Could you please guide me on how to access the algorithm and witness its effects? Your insights or any relevant documentation would be greatly appreciated

aavek commented 1 year ago

Hi! Thank you very much for your kind words.

So, currently Aeolus-Ocean (AO) has been made public as a demonstrator. Effectively, this means that you cannot access the underlying code for the Agents or alter the hyperparameters that were used in the simulation as these were prepared prior to the compilation of the binary.

Having said this, I could point you in a couple directions. See our paper, for how you yourself could try build an environment like AO, since you want to test on a platform like this. In order to test various reinforcement learning algorithms as you mentioned, I would recommend you visit the Unity ML-Agents repository - they really have some excellent resources to go through and will give you a clearer idea of how to apply them in Unity itself.

Also, it is far better, and fruitful, if you try playing around with RL algo's in a much simpler environment than AO. Try using one of the ML-Agent environments and follow their setup carefully, up to the point where you yourself will have to go into a script or two and alter the actions/gameplay/mechanics to something that suits your needs.

I hope this helped somewhat.