Simulating and Learning System Behaviour
Botie can learn from simulated experiences, but we first need to know what that simulator is doing. Inspect numerical time steps, compare policies on shared scenarios, fit unknown physical parameters and test conditions that the original model did not handle.
Learning from a numerical world
A simulation chooses how to approximate movement and which conditions to generate. Work through those choices before using the examples to judge a learned policy.
- 01Ready to exploreThe Simulator Invented an Oscillation
Follow Euler integration and distinguish a numerical timestep failure from physical behaviour.
- 02Ready to exploreGive Two Policies the Same Simulated Shift
Compare policies on shared random scenarios and report uncertainty in their paired differences.
Checking a simulator against observations
Some relationships in our simulator are unknown. Fit them from measurements, then try the fitted model under conditions it has not already seen.
- 03Ready to exploreLearn the Unknown Part of a Physical Model
Estimate a system gain from measured commands and responses, then inspect identifiability and model mismatch.
- 04Ready to exploreThe Policy Works in One Simulator Setting
Evaluate a learned policy across changed physical conditions and keep the worst tested condition visible.