No details provided now.

1. Always assert when you are not sure;

2. Range Check Assertion: $(a,b), [a,b], [a,b), (a,b]$;

3. When initialize array: argmin with np.full, argmax with np.zeros.

• try fastmath option in @njit (no performance difference)
• commit a main_one_shot function, with reduced and formatted output
  • move all related traces under same folder
  • recording somehow per-stage
  • complete main_one_shot function
• add static policy replacement function
• test policy replacement (with 50 submission)
  • running on vps and get 50 results
• semi-analytical average cost calculation
  • one-step/n-step policy improvement for any stage (n < STAGE_EVAL)
  • Possible: enhance one evaluation with multi-step policy improvement's return?
• finish two simple analysis
  • reinforcement learning
  • optimized baseline policy (aware of start)
• touch a plot-traces2.py with new plot function
  • calculate the record data and display