3d42de7b12
Previously the game ran indefinitely and relied entirely on max_turns_per_episode to end an episode. Add start/food energy values so an episode ends on its own when energy runs out, and mention retro-gamer plot in the lab's training log instructions.
1.7 KiB
Forager Training Log
Document each training attempt below. For each attempt, write your hypothesis before you run the experiment, then fill in the evidence and analysis after.
Use retro-gamer info runs/forager/ to see a summary of your run,
cat runs/forager/training.log to see the full log, and
retro-gamer plot runs/forager/ -o runs/forager/training.png to graph it.
Attempt 1
Hypothesis
Before training, predict what will happen with the default configuration. Will the agent learn to find the food? How quickly? What might go wrong?
Your prediction:
Configuration
Copy the relevant sections of runs/forager/config.toml here.
Evidence
Paste the first and last few lines of your training log, and any interesting moments in between.
Analysis
What happened? How do the numbers — avg_reward, avg_steps, epsilon, avg_loss — tell the story of what the agent learned? Did the result match your prediction?
Attempt 2
Hypothesis
Based on what you observed in Attempt 1, what will you change and why? Predict the outcome.
Configuration
Evidence
Analysis
Attempt 3 (if needed)
Hypothesis
Configuration
Evidence
Analysis
Final analysis
Which attempt produced the best-trained agent? Run retro-gamer play on your
best run's checkpoints and describe what the agent does.
Your answer:
Compare two of your attempts. What changed between them, and how did that change affect the training curve?
Your answer:
If you had more time, what would you try next to improve the agent further? Refer to specific hyperparameters or configuration options.
Your answer: