Another idea came to me when playing Tetris myself. An obvious one too. Give the agent a jump-start in placing pieces by recognising the best places for the piece to go. Do this by matching contours to the piece’s structure. So, for example, an S-piece would sit best when horizontal on a vector {0, 1} going by the contour structure. When vertical a vector of {-1} would work.
Going by this, an I-piece would sit best anywhere when vertical, but best in chasms or the edge. The system may need reworking for all this new structure.
This idea seems better as the agent would be less random and go by guided exploration to maximise reward gained while exploring.