Bin-Packing Agents
First, Best, Worst, and Next Fit strategies reframed as online row packing — Best Fit within 0.5% of El-Tetris.
IEEE Industry Track · Experience Report
Engineering a Pluggable Tetris AI Workbench
Bin-packing heuristics, Double DQN with ARC episodic replay, and an honest CPU/GPU acceleration study — documented in a peer-style industry paper by Sapana Micro Software.
Read & download the paperAuthor
Shyamal Suhana Chandra
Chief Engineer (Manager)
Sapana Micro Software
Pittsburg, KS 66762, USA
6+
Agent families
398
Lines cleared (El-Tetris)
~34
Placement candidates
9× CPU
Accelerate speedup
System at a Glance
A unified Swift 6 substrate hosts heuristic, search, exact-DP, bin-packing, and deep-RL agents behind one protocol — making benchmarks fair and defects visible across every algorithm.
TetrisCore
Board · Game · Placement
TetrisAI
Heuristics · Search · DDQN
TUI
Ncurses Renderer
TetrisApp CLI
Play · Train · Benchmark
Paper Highlights
The industry-track report covers architecture, empirical benchmarks, a root-cause case study, and reproducible compute measurements.
Double DQN + ARC Replay
Afterstate value learning with Double Q targets and Adaptive Replacement Cache episodic retention.
Exact Offline DP
Memoized optimal solver over short horizons, inspired by Demaine et al. offline Tetris complexity results.
Honest Acceleration
CPU Accelerate beats GPU at every batch size tested; GPU exposed as opt-in with throttling.
Shared-Code Debugging
A placement-enumeration defect masked all agents at zero lines — fixed by restoring full rotation × column search.
Terminal UI
Bounds-aware ncurses renderer with human play, AI watch mode, and active learning from manual sessions.
Read the Paper
The full IEEE industry-track PDF is published here. Preview inline or download for offline reading. Source code in the private repository is not distributed — only this paper is public.
Direct link: paper.pdf