The Champion: a Neural Network Trained for the 9x8 Board
Every level in Line4 9x8 is played by the Champion - not a scripted bot with canned openings, but a neural network trained specifically for the 9x8 four-in-a-row board. It reads the position through two learned heads: a policy that suggests promising columns, and a value that judges who stands better. What changes between levels is how much of that judgement the Champion is allowed to use - and, at the bottom of the ladder, how often it is allowed to be wrong.
How a move is chosen
- Levels 1 to 3 make real mistakes, on purpose. They skip tactics they can see and sometimes blunder outright, because an opponent you can never beat is not an opponent - it is a wall. Level 1 loses well over half its games to a player who simply takes wins and blocks threats.
- Tactical rails, from level 4 up. From level 4 the Champion always takes an immediate win and always blocks yours - it will never miss a one-move tactic again.
- Learned judgement. The middle levels blend the network's policy with a safe local heuristic, sampling more loosely the lower you go.
- Opening book. The upper levels open from a book of 7,248 verified positions - cached results of the engine's own search.
- Value-guided search. Levels 7 and 8 evaluate every legal reply with the network and pick the move that leaves you worst off, using the policy as a tiebreak.
- Monte Carlo tree search. The two Max levels - available in the Android app - spend a 64- and a 256-simulation neural search on every move.
A benchmarked ten-level ladder
The levels are calibrated against each other, not guessed: 96 head-to-head games per adjacent pair under a fixed protocol, seats swapped, and the ladder came out strictly monotonic - every level scores between 62% and 76% against the one below it, which works out at 84 to 198 Elo per step. Level 4 is the anchor at 1000.
| Level | Tier | Elo | How it plays |
|---|---|---|---|
| 1 | Easy | 629 | Learning the board - misses tactics, blunders often |
| 2 | Easy | 732 | Casual play - still misses things |
| 3 | Easy | 916 | Attentive play - rarely blunders |
| 4 | Easy | 1000 | Tactical play - never misses a one-move win or block |
| 5 | Medium | 1148 | Sharper tactics |
| 6 | Medium | 1259 | Policy-guided play - the network's instincts |
| 7 | Strong | 1457 | Sampled value search |
| 8 | Strong | 1584 | Value-guided search plus the opening book |
| 9 | Max | 1750 | Neural search, 64 simulations - Android app only |
| 10 | Max | 1865 | Neural search, 256 simulations - Android app only |
A benchmark is evidence of strength on its own terms, not a theorem about every position. For a much deeper measurement - 1,395 seeded games, every opening mapped, downloadable data - see the 9x8 self-play study, which was run on the four-level engine that levels 4, 6, 8 and 10 are inherited from.
The same engine, in your browser
The web version runs the exact same neural network as the Android app, compiled to WebAssembly and executed locally in your browser - your games never leave your machine. Levels 1 to 8 are playable at line4-9x8.com; the two Max levels play in the free Android app.
Beat it, then train
Start with the rules and strategy basics, win a few at level 1 or 2, and climb. When level 7 starts punishing your openings, sharpen your eye with the Win-in-2 Tactics puzzles.