Kubisme (Cubism) is an AI trying to play Tetris.
Parameters (78th generation) without I-potential.
Parameters (60th generation) without I-potential.
Removed I-potential detection.
Rolled back the 3-ply parameters.
Parameters based on 3-ply simulations and a bugfix in lock row.
Taking opponent first filled into account did not pay off. So removed.
SingleEmpties not longer commented out (ouch!) New parameters. Fix in finding skips, and finding paths on ply 2 and deeper.
Overflow fix on opponent.
Take first filled of the opponent into account. Removed evaluation of tunnels and other unreachable potential. Is seems not to work.
New way of handling curves in parameter generation.
Always check on unreachable holes. Check paths in endgame. Added evaluation for to penalize unreachable rows with potential.
Take height of unreachables into account when valuation unreachables.
Rolled 3ply stuff back too.
Rolled back to Version 65 with 'Improved speed of path finding enormously' of Version 66.
New parameters fine-tuned on 3 ply.
New parameters and exclude delta score for more than 3 ply.
Increased opponent analysis to 3 ply. Only search for I-potential with 'deep tunnels'. Take first filled of opponent into account for the score of a field. Improved speed of path finding enormously.
Extended curved parameters.
Introduced curved parameters, and single T-spin detection. Score 0.67.
Divide by zero exception bug fix for winning on ply 2 and 1, and losing on the next turn.
Fixes for bugs with getting T-spin bonus (resulting in way to high scores). Added all 7 blocks for ply 3. Score 0.60.
Parameters with a score 0.74.
With parameters of version 56.
Fixed endgame evaluation, improved T-spin detection. New parameters, and unreachables count as two holes per line.
For now, only 2 ply deep. Tweak in factor for unreachables.
Hopefully improved parameters.
Split parameters in default and endgame, and added second ply for opponent evaluation.
Rolled back the parameter changes, as they performed shit.
Fixed a painful block path finding bug. Updated parameters because of this bug.
Added penalties for single empties.
Send the last move that was found, not the be losing directly, and an update of the logging.
Re-introduced garbage in the evaluation, using the worst outcome of the current block, and detecting potential win on the current block.
Just new parameters.
Added losing-changes for row 1/4, and keep that in the evaluation for depth 1.
Re-introduced combo-potential. Removed score for empty cells (not for empty rows). Use 'locks' instead of 'garbage' to simulate less space through time (due to actions of the opponent). Both faster and less random.
Removed unreachable distance detection, changed branching factor. Added detection of multiple row clearance potential.
Removed wall score, added distance to unreachable hole.
Totally re-written field evaluator.
Fixed a Skip bug. (Skip instead of Skips,drop)
Removed opponent logic, simplified the evaluator.
Applied new rules, without taking the 'skip' move into account.
Redesign of time management, and fix of evaluation of row difference.
Bug-fixes from live resolved.
Changed branching, used a fixed number of blocks for ply 3 and deeper.
Due to disk space issues on the server, this compilation failed.
Radically change in evaluation the position. Move focus on the position relative to the opponent.
Tweaked parameters based on one line of garbage per 4 instead of 6 points.
Fix for positions of S and Z block after rotation left and right.
Detection for unreachable lines (mainly garbage) and the reachables.
Bug fix for paths of I,S, and Z blocks. Tweaked detection for T-Spin.
Fix in path generation for long (30+) paths. Added T-Spin detection and parameters.
Fix in path finding. added T-spin point. Changed evaluation; check for free rows, instead of free cells.
Last blockade detection and path finding for filling holes.
Disabled combo bonus for rows of 9.
Tweaked parameters based on the simulation fix.
Yet another time consumption tweak.
Fix in time consumption.
Tweaked branching factor based on available blocks, potentially leading to two ply more.
Re-factoring if the block object. Different implementation handle block type specific logic, including a test on the availability of the path.
New parameters and a block node lock row fix.
Max depth to 5, and higher branching, to prevent time-outs.
Fix in add garbage rows.
Bug fixes and testing added. New live data needed.
Clearing fix, extended logging.
Disabled filling (for now), implemented new rules.
Fix in starting position, first taken in to account in Version 16.
Further tweaked the parameters. Reduced the branching factor even further. On top of that, Introduced path finding for filling (reachable, duh) holes.
Introduced a rudimental combo potential detection, and tweaked (increased) the branching factor a bit. Based on a stronger virtual opponent and the compo potential, updated parameters as well.
Fixed the determination of blockades, and of the the contacts with the right wall. Furthermore, the parameters were updated. On top of that, the search depth is extended. In practice, it searches up to 6 ply.
Failed to compile on TheAIGames.com due to some not supported reflection code?!
Added block per row count specific weights.
Changed the genetics for optimizing evaluation parameters. With improved parameters and more checked variations.
Updated evaluation parameters. (Not that successful).
Fixed the swapped L rotate left and right.
Fixed rotate right bug (was move right).
Fixed a lot of bugs, enabled evaluation of one 'unknown' block as well.
First attempt. Local runs (without competition) run 75 points on average before dying.