1
0.1
0.01
one half
time, 0 to 68 (step size × steps)
policy gradient
demonstration
starting probability ε
0.01
reward gap Δ
1.0
Time to reach one half: policy gradient
54.1
, demonstration
2.8
, a ratio of
19×
. Policy gradient grows like 1/(
ε
Δ); the demonstration like log(1/
ε
).