-
Notifications
You must be signed in to change notification settings - Fork 15
b2developer/SpidermanPPO
Folders and files
| Name | Name | Last commit message | Last commit date | |
|---|---|---|---|---|
Repository files navigation
Load the Unity scene to start
Press the Spacebar to start the simulation
The AI uses compute shaders to speed up matrix calculations, so if you don't have a GPU
it will probably cause some issues / significant speed losses.
PROJECT SETTINGS
fixed timestep = 0.01 - the AI runs 100 times per second (mainly for stability of physics)
physics.autosimulation = false - the AI runs the physics
BOOTSTRAP (unity object)
----------------------------
the networks are stored in 4 text files: mu, sigma, value and running. mu and sigma store
the actor's neural network, while value stores the critic's neural network. Running
stores some moving averages that make the training more efficient
TESTING - true (shows the AI running), false (actually trains the AI)
LOAD_ID - the name of the network you want to load. If it doesn't exist
the app will create one
FOLDER_PATH - the path of the folder containing your neural networks
RUN_ID - the name of the AI, when new training sessions finish, the AI will be saved
under this name with it's performance statistics along side it
COUNTER - keeps track of the current training session, you'll need to set this manually
to the highest number if you start the application multiple times
NEXT_CHECKPOINT - milestone for timesteps, after this number is reached the AI is saved
CHECKPOINT_INTERVAL - how far apart should automatic saves be?
i've included several AI shown in my video that you can load up using the LOAD_ID
----------------------------
PHYSICS MANAGER (unity object)
----------------------------
the simulation is managed by this object
MODE - REALTIME (runs at normal speed), FAST (runs as fast as possible)
ENVIRONMENTS - a list of all the active environments, you can add more than one if you like
STEPS - how many steps does the FAST mode do at a time?
----------------------------
About
Deep Reinforcement Learning (PPO) applied to a Spiderman Swinging Environment
Resources
Stars
Watchers
Forks
Releases
No releases published
Packages 0
No packages published