I have started using curiosity in the config file and the agent actually leaves the first room but for some reason all of the agents that are training go to the same room and never leave even though there are areas that have not been explored. Is there something that I am missing? currently I have cubes in each section and in order for the level to reset they have to colllect all six in the area. I thought this would be a good way to train the ai to move around the map. The end goal is to have the AI fight each other like in an FPS. If anyone could give me suggestions on the different stages of teaching that would be appreciated as I feel like I am not going in the right direction. I have attached the level I have created as well as the config file that I am using. I am new to using ML Agents so any advice or suggestions on how to train an AI to play an FPS would be appreciated.
Hi,
I’m not sure but do you have observations other than the raycasts? Like the position of your agent?
The second thing can be to increase the strength of curiosity.
Oh by the way don’t forget to indent reward_signals in your config file.
Hope it helps
I have a public bool called use vector obs but im not sure that I am actually using it, thanks I’ll have a look into it!
I am also confused about vector observations. What are they actually used for because i thought the ray perception sensor scripts replaced using vector obs.
As an example, vector observations could be relative x,y,z coordinates of or a vector towards an entity in a game. This is distinct from raycasts which is just a relative distance along a ray to an object with a specific tag.