Policy gradient cart (REINFORCE)
Settings
Settings
Learning rate
0.0030
Discount γ
0.980
Entropy bonus β
0.010
Reward
+1 catch · −d/10 miss
+1 catch · −1 miss
Speed
1.0×
Game size
100%
Baseline b (subtract the average reward so far)
Fast-forward (skip animation)
Show where the ball will land
Game
Episode 1
← Left
Coast
Right →
θ ← θ + α Σ
t
γ
T−1−t
[
R − b
] ∇
θ
log π(a
t
| s
t
)
Step
Run continuously
Reset