A Cadence brain plays a game it was never told the rules of

Ember Field

Cadence brainone continuous life
What it sees right now
Plain neural net (MLP, Q-learning)same embers
What it sees right now

A whole life, 45,000 moments

What you're watching

Cadence is an open-source brain library from Pragma Research. Instead of being trained offline and then frozen, a Cadence brain lives: every moment it settles on what it sees, acts, and adjusts itself from what that action brought. It keeps an associative memory of what paid in which situation and votes with it, and it keeps learning for as long as life pays less than it wants, which in this game is always. Here one brain is born into a small field with no rules explained, only pay. Each half of the replay is one continuous life; the clips are cut from it at six points.

Field
9 columns by 7 rows. The lantern walks the bottom row.
Moves
Left, stay or right, once per moment.
Pay
+1 for catching the food colour, −1 for the sting colour, 0 for a miss.
Seasons
Amber is food until moment 15,000, blue until 30,000, then amber again. Nothing announces the turn.
Senses
16 numbers: for 3 columns either side and its own, how close the lowest amber and lowest blue ember are, plus a wall on each side. The strip under each field draws them.

Across 12 lives

Limits

  • One small game I designed, 12 seeds per player. The Cadence settings (memory voting on, a need tied to the spread of its pay) were picked in a short sweep on other seeds; the MLP got no tuning.
  • The plain net is the better player overall. The Cadence brain learns to dodge reliably but only partly learns to chase food, and no setting I tried in a short sweep (critic rate, eligibility, temperature, memory rate, a second region) closed that gap.
  • The showcase life is the median Cadence seed by final-season score, not the best one.
  • A replay of recorded lives. The brains learned in Python on a desktop, not in your browser.