TrainingMonitor

Logs episode statistics and plots training curves. Automatically used by Runner, but can also be used standalone.

from tinyrl import TrainingMonitor

monitor = TrainingMonitor(window=50)

Constructor

TrainingMonitor(window=50)

Args:

• window — rolling average window size for smoothing plots

Methods

log(reward, length, entropy=None)

Record one episode's statistics.

Args:

• reward — total episode reward
• length — number of steps in the episode
• entropy — mean policy entropy over the episode (optional)

plot()

Display training curves side by side:

1. Episode Reward
2. Episode Length
3. Policy Entropy — only shown if entropy data was logged

Each plot shows the raw data (transparent) and a rolling average (solid line).

Attributes

Attribute	Type	Description
rewards	list[float]	All logged episode rewards
lengths	list[int]	All logged episode lengths
entropies	list[float]	Logged mean entropies (only for episodes that provided entropy)

Standalone usage

from tinyrl import TrainingMonitor

monitor = TrainingMonitor()

for episode in range(500):
    # ... your training loop ...
    monitor.log(total_reward, steps, entropy)  # entropy is optional

monitor.plot()