Continuous Fitted Value Iteration based on closed-form solution of Hamilton-Jacobi-Bellman equation for affine systems, implemented in JAX. This method was first used in this paper and tested on a number of classic contorl problems like cartpole or pendulum swig-up tasks.
- Extensibility for custom environment definition
- Linearization of forward dynamics and reward functions
- A number of different solvers for bellman backup optimization
- Various integrators for forward dynamics
- Various data collection and storage methods
- Extensive configurability
- Tensorboard and file logging
For now:
python main.py --config-file=./config/Pendulum.yaml
tensorboard --logdir=./logs/<log_dir>