Run domains in terms of sample complexity as well as computation time
Comparisons to other planners
- How does performance of discrete planners degrade as dimensionality increases
- How does cost of discrete planners increase as dimensionality increases (computational, memory)
- Tolerance of noise
- Domains (beat UCT, maybe OLOP, FS3):
- Double Integrator
- Inverted pendulum
- Bicycle, need to introduce a new (non-binary) reward
So if we can beat discrete planners, how close are we to optimal performance?
- Double Integrator (can we beat LQR?)
- HillCar (even though its bang-bang, and requires “good” exploration, how close do we get to optimal?)
Domains that are hard enough that any result is good
- Articulated planar arm
- Down Stairs
- Down collapsing stairs
- Responding to shoves
- Model Building?
- Using trajectory libraries?
- Reinitializing search based on where previous search converged to