Win rate, rolling over 500 runs (whole training)
Average wave, rolling over 500 runs (whole training)
Reward per run (last 2,000 runs)
Wave reached per run (last 2,000 runs)
Training loss
Average Q (expected future reward)
Exploration (random decisions)
Average wave per 100 runs
Newest first, over the whole training (from summaries of 100 runs each, so the newest row can still be filling up). The trend tiles use the last runs one by one. ± is the noise of the average (standard error); a change is colored only when it is bigger than twice the noise. With several danger levels mixed the average also moves with the mix: pick one level to compare like with like.
Win rate, last 500 runs
Stats it values
Level-up options
Each upgrade by its effect and tier, compared with the average option offered.
Shop: buy instead of leaving
Q(buy item) − Q(go to next wave): positive = it would rather buy it than leave the shop. Offered = times it was in a shop the agent could afford; bought = times it was bought (this session).
Results per starting weapon
The starting weapon is random; it should learn to win with all of them.
Where runs die
Strong vs. weak runs: the build after wave 8
Average of each group (last 400 runs). Big differences show what makes runs go far.
Results per build
What the weapons scale with at the end of the run (60%+ of the scaling; else "mixed").
Materials left when pressing Go
Average per run, per 100 runs (whole training). Good players spend nearly everything.
Actions per run
Average per run, per 100 runs (whole training).
Memory per game copy
Private memory (MB) over the time the copy has been running. Rising lines = a leak.