Researcher hand-codes MLP weights for Breakout on GitHub

"The first layer computes four landing-position errors and two direction indicators." That is how a short GitHub note opens, setting out numeric weights for a small network that plays what the filepath labels "breakout."

The author describes each unit's role in a few flat sentences. "Descending predictions include wall reflections; ascending predictions anticipate the earliest possible brick bounce and revise as the ball climbs." Early in a rally the policy favors "preserving horizontal direction to sweep through bricks"; near the bottom of the screen, "either valid catch column is accepted." A second layer "gates left/right decisions, choosing no-op when positioned correctly."

Below that sits a block of explicit numeric assignments: W1, W2, W3 matrices indexed by ball row, with the first four output units commented as "descending-right, descending-left, ascending-right, ascending-left." No training run, no learning curve, no score. The file was moving through a couple of the researchers on our radar this week.