Case Study: Five Legends by Search

01Nine games, three kinds of bot

You've reached Gold in a CodinGame game and the boss above you won't move. The next version could be a deeper search, better rules, or a network trained by self-play, and each one costs days. The author's lab made that choice nine times in September and October 2026 and recorded every submission. As of October 30, 2026, eight of the nine games had reached Legend, the top league. Three got there with self-play networks. The other five got there with search bots or rule-based ones, whose decisions follow rules written directly as code, with no learning and no lookahead, each built from the game's published post-mortems.^[1]

A boss is a bot CodinGame copies into every league except the top one. From Bronze up you become eligible for promotion when your score, a skill rating computed from your ranked games, reaches the boss's once all of them are played.^[2] The widget below measures exactly that gap: each bot's final score in Gold minus the boss's.

Only one of the nine games, Back Track King, was a live contest. The other eight are permanent multiplayer puzzles: bot games that stay open all year, most of them left behind by earlier contests, with every league already open and a field that changes slowly. In a live contest the top leagues open on dates set in advance and every player is improving at the same time.^[2] A climb on a permanent puzzle says little about a contest placing, so every rank on this page names which one it was.

The kind of bot followed the shape of the turn. The three turn-based games, where a bot can simulate every move exactly and fast enough to look several moves ahead, went to search. The two games where good play is a plan more than a lookahead, a racing line in Mad Pod Racing and an economy in Troll Farm, went to rule-based strategies. Networks reached Legend in Back Track King, where one turn combines several placements, in Fantastic Bits, and in Spring Challenge 2022 after search had stalled. That's nine games in one lab, so treat the pattern as a first guess for your next game and then measure it.

02Ultimate Tic-Tac-Toe: when nobody can score a position

Drew on: the game's forum thread (Magus, jacek, JFB, RoboStac) and Tom Alard's 2026 write-up.

In Ultimate Tic-Tac-Toe, nine tic-tac-toe boards sit inside a big one, and the cell you play decides which small board your opponent must play in next. A turn can offer up to 81 legal moves, and nobody has a good formula for who's ahead in the middle of a game. Tom Alard, 2nd on the leaderboard when he wrote it up in 2026, puts it plainly: "A classic minimax search that greedily tries to maximize some naive heuristic will play very poorly."^[3] When you can't score a position by hand, let random games score it.

That's what Monte Carlo tree search does: it estimates a position by playing many random games to the end from it, and spends more of them on the moves that look best so far. The game's forum thread reads like a checklist for it. Magus reported reaching Legend in 2018 "with 25k rollouts at the second turn", keeping his tree between turns.^[4]

In a later post, jacek listed what got his Java bot to Legend: "mcts solver; fine-tuning exploration constant (it's generally much less than typical sqrt(2)); tree reuse; 1-ply search win check during expansion and during playouts". He added later that "The best opening is center, almost all in legend use it."^[4] JFB's 2018 post is the counter-example: he trained a small network on about 2,000 replays, played it without search, and reported "It finish as last in Gold."^[4]

Two habits carry over. Tune the exploration constant c, the weight UCT gives to rarely tried moves, on your own reward scale instead of copying a number: with rewards of 0, ½ and 1 the forum's values and the lab's sit well under √2, yet a 0.3 to 0.5 range taken from another bot's tests never beat 0.6 here. And measure each addition against the bot without it. Tree reuse, keeping last turn's tree under the moves actually played, was worth .620 in this bot, while RoboStac, answering a newcomer on the same thread, called it "a very small increase in bot strength" and something to leave until MCTS works.^[4]

Random playouts have a ceiling. Tom Alard estimates that "excellent performance engineering" takes random-rollout MCTS to "about top 50", and the top bots he describes put a small neural network inside a search, alpha-beta in one case and an MCTS variant in his own.^[3] Monte Carlo Tree Search builds the bot above, and Networks Inside Search covers the step after it.

03Tron Battle: score the territory, then search

Drew on: a1k0n's and Janzert's write-ups of the 2010 Google AI Challenge, dhartmei's notes on chambers, and the game's forum thread (dolmen1234, Xavier_F).

In Tron Battle, light cycles leave a wall behind them on a 30×20 grid and the last one moving wins. A cycle has at most three useful moves a turn, so searching is cheap and knowing who's winning is the hard part. The winner of Google's 2010 AI Challenge, a1k0n, played a version of the game where both players moved at once, and wrote that "a better evaluation heuristic will always beat deeper minimax searches".^[6][7] That's his opinion, and the lab's Legend bot was built on it.

His evaluation, which he credits to the contest's forums, is Voronoi territory: for each free cell, find which player reaches it first, then count your cells minus theirs.^[6] He added two refinements. Counting the open edges inside each territory as well as its cells "led to automatic wall-hugging behavior when all else was equal". And he corrected for chambers: when part of a territory hangs off a single cell, an articulation point, a player who goes through it can't come back, so only the side it can actually fill should count.^[6]

Janzert, 5th that year, built the same pieces: "crAB is an alpha-beta searcher using iterative deepening, a transpostion table and bitboard representation. With Voronoi territory and 2-connected chambers used for evaluation."^[8][7] Once the players are walled off from each other, the game becomes a filling puzzle, and dhartmei's notes solve it chamber by chamber.^[9]

CodinGame's version adds players: of the lab's 1,770 Gold ranking battles, 60 had two cycles, 366 had three and 1,344 had four.^[1] On the game's forum, dolmen1234 reported that a paranoid search over all players "seemed to improve my results and get me to legend", while Xavier_F reached Legend in Python with a one-ply heuristic and no search at all.^[10] Paranoid search is one reported route among several.

Put your first days into the evaluation and the endgame, then let alpha-beta use whatever time is left in the turn. A deeper search over a territory count that ignores chambers steers into a pocket it can't fill. Evaluation Functions builds Voronoi territory, and Minimax and Alpha-Beta adds paranoid search and the endgame.

04The Great Escape: search only the moves that matter

Drew on: the 2015 contest's strategy thread (Recar, 1st; Romka, 3rd; dimkadimon) and CodinGame's final report.

In The Great Escape, two or three players race across a 9×9 board to the far side, and on any turn a player can place a wall instead of stepping, as long as every player keeps a path to its goal. A wall can go in up to 128 places, so a full-width search grows fast. Romka, 3rd in the 2015 contest, decided early "that minimax is not the way to go because of large amount of possible wall placements on each turn".^[11][12]

The winners cut the move list down. Recar, 1st, started with "mini-max with alpha beta pruning", switched to a list of rules checked in order, and kept one narrow search in which the "Enemy can only select one of decreasing his path turns" and walls go only where they block the opponent's next moves.^[11] Romka searched walls to depth 5, "assuming that opponent won't put any walls and will just move towards his goal".^[11] In three-player games both picked a single rival; Recar's rule was "Usually our rival is next player."^[11] And dimkadimon described a fixed opening, credited to Laochao, that made "a nice passage for himself around the edge" and "usually won against the top players at the time", until players started countering it.^[11]

When the move list is huge, prune it to the moves the post-mortems say change the outcome, here walls that lengthen a path, and check the opponent model against real games before trusting it: both winners assumed the next player was the rival, and the lab's recorded games put 85% of walls on the next player or the leader. Expect a fixed opening to stop working once others have seen it, and treat it as a head start for the search.

05Mad Pod Racing: one pod races, one pod blocks

Drew on: write-ups from the 2016 Coders Strike Back contest: Hastings (4th), pb4608 (3rd) and Magus's guide.

In Mad Pod Racing, pods race through a circuit of checkpoints, and in the full game each player steers two of them; the first pod to finish the laps wins for its team. The game ran as a contest in 2016 under its previous name, Coders Strike Back, so that's where its post-mortems are.^[13] Steering one pod well is the first half. Hastings, 4th in that contest, aimed at "nextCheckPointPosition - 4 * velocity" to cancel the pod's drift, thrusting only when roughly facing that point, and wrote "I used 4 because it performed better than 3 or 5."^[14]

The second half is the other pod. Magus's guide gives each player a runner and a hunter, re-chosen at the start of every turn from the pods' progress.^[15] The 3rd-place finisher, pb4608, opens every turn the same way: "Define all pods' roles: my runner, my blocker, the opponent's runner and the opponent's blocker."^[16] Hastings's heuristic bot made the split by asking whether a pod was "the furthest along on its team".^[14]

Promotion didn't end the work. In Legend the same bot finished its ranking in the lower half of the table while still capping its thrust at Gold's 100, where Legend accepts up to 200. A copy with Legend's thrust won 16 of 20 paired live games against Legend players, where the Gold copy won 6 of the same 20.^[1][5] CodinGame's league documentation warns that a new league can bring new rules, so "read the statement again after a promotion".^[2] The Legend copy was submitted on October 30; the records read that evening hold no final ranking for it.

06Troll Farm: write the economy down

Drew on: delineate (1st), yamo (3rd), gpoussel's notes on a hand-written Legend bot, and the game authors' Legend statistics.

The Spring Challenge 2026 game, Troll Farm, is an economy. Trolls harvest fruit, chop trees and mine iron, and you spend what they carry home on training more trolls, so the team gathers faster. A troll trained now costs resources today and pays them back over the turns that remain, so whether to buy one depends on how much of the game is left. Plain self-play struggled with exactly that decision in this game, the winner's included.

The contest's winner, delineate, used a network, and his write-up explains the trouble. His first version "could never learn to mine iron, so it never ventured into higher tech units", because "The benefits to mining iron are extremely far in the future."^[17] He fixed it with a curriculum: first a network rewarded for building one specified troll design, then one that chooses among designs, then a fine-tune on the real final score.^[17]

The players without a network wrote the economy down as rules instead. The 3rd-place finisher, yamo, rates each tree by wood per turn, what a troll can carry home divided by the whole round trip: "value = min(treeSize, carryRemaining) / (travelTurns + chopTurns + returnTurns)".^[18] Notes by gpoussel on a hand-written bot promoted to Legend in the post-contest arena use the same rule for every job: "Each troll picks the job with the most points per turn".^[19] The authors' statistics for the Legend league show two economies at the top: yamo ended every game with two trolls, while delineate and norxondor_gorgonax, 2nd, usually built three or more.^[20]

A discount factor is the other common way a learner misses an investment, and it's the first thing to check in your own trainer. A reinforcement learner usually counts a reward t turns away at γ^t of its value, where γ (gamma), the discount factor, is a number just below 1, so income far in the future shrinks toward nothing.

The widget plots this sum for every turn of the game. It's what one more worker is worth to a discounted learner:

In words: add up the worker's income for every turn left, shrinking each turn's share by another factor of γ, then subtract the price. However many turns are left, the sum stays below income / (1 − γ), so when that is less than the price, no purchase ever looks worth it.

Write an economy as rules when you can state it: rate each job by points per turn, and buy the next producer when its output over the turns left beats its price. If you want a network to learn it, give the network the situations to learn from, as delineate's curriculum did, and check that its discount factor can see as far as a purchase takes to pay back.

07Where search stalled

Two games broke the pattern. In both, the lab's search bots climbed near the top of Gold and then stopped short of the boss, round after round. A retrained network finished one of them; the other is still in Gold.

Spring Challenge 2022: a network finished it

Drew on: bowwowforeach (1st), blasterpoard (2nd) and Nanaeda (3rd).

In Spring Challenge 2022, listed today as Spider Attack, each player's three heroes guard a base against monsters walking in from the edges of the map, and a base that loses three health points loses. Heroes earn mana by hitting monsters and spend it on three spells: WIND pushes units away, SHIELD protects a unit, and CONTROL sends a unit where you choose. Games where neither base falls go to the player with more "wild mana", the mana gathered away from its own base.^[21]

The top three write-ups describe three different bots. The winner, bowwowforeach, ran two rule-based defenders and one attacker chosen by a 15-turn search ("防衛2人 : ルールベース", two defenders by rules, and "15ターンのchokudaiサーチ", a 15-turn chokudai search, which he describes as a depth-first relative of beam search).^[22][21] The runner-up, blasterpoard, used "no search" and scored several thousand candidate points with an evaluation function.^[21] Nanaeda, 3rd, defended with all three heroes until turn 120 and then attacked with shielded monsters, a plan he chose because most top bots attacked early.^[23]

When round after round of search changes lands inside the noise of a fair panel, another search tweak is unlikely to break through. Look at what the bot is measured and trained against instead. Here the local copies of the shield attackers didn't predict games against the real ones, and the network moved once its reward, its actions and its opponents matched the Gold meta.

A common question at this stage: the network has stopped improving in Gold, so what should change? Spring Challenge 2022's records show five causes. The network drifted into a habit of farming the tie-break. It never cast SHIELD, an action its cloned teacher never used. Greedy play stalled where sampling kept moving: one champion scored .50 picking its most likely move and .85 sampling, on the same maps. Its heroes decided independently and duplicated each other's work: two of them controlled the same monster on 254 turns of 52 lost battles. And no training opponent played like the Gold meta. Each fix changed what the network sees, does or plays against.^[5] Behavior Cloning covers the forgotten action, Neural Networks for Game Bots has units decide in order, and Self-Play Reinforcement Learning covers drift, sampling and opponents.

Game of Drones: still in Gold

Drew on: Loujo (1st), god (2nd) and Gangrene (3rd) from the 2014 contest, and the game's forum thread (tarriel).

In Game of Drones, two to four players move drones over a map of zones. Each zone you hold scores a point a turn, and you lose it when an opponent brings more drones into it than you have.^[24] The 2014 winners all planned in turns of travel. Loujo, 1st, listed for every zone the future turns at which drones would be needed to take or defend it, sorted those objectives by how soon and how cheap they were, and assigned drones in that order.^[24] The runner-up, god, kept his commitments loose: instead of "From now on I'm going to charge into this zone with 3 drones!", his objectives read "I want to be able to have 3 drones there in 5 turns".^[25] With three or four players, god and Gangrene, 3rd, both settled on an edge base, and Gangrene limited his drones to "a zone at an extremity (zone A), but also its two closest neighbors (zones B and C)".^[25][26] Loujo also noted why testing a bot against a near-copy of itself teaches nothing in this game: "the score remains 0-0 at the end".^[24]

When you're first in Gold and still below the boss, split your ranked results by opponent and by player count before writing more code, and work on the games you're actually losing. On the game's forum, tarriel gives the same advice: "reread the post competition thread", because "there are usually commonalities between the best bots and everyone else".^[27] Here the commonality was the edge base, and it moved the score without yet passing the boss.

08What to reuse in your next game

The five Legend bots on this page share no code, but each one went through the same checks before it was submitted, and those checks transfer to any game.

"Local" is the bot's points against the author's earlier bot for the same game; "live tests" are test games against the Gold boss before submission. Points count a draw as half a win, and the brackets are 95% Wilson intervals. Every local match used fixed validation maps, with both seats. Troll Farm's Gold final was read with 97% of its ranking games played.^[5][1]

1. Read every write-up and list its claims. Write each one as a change you could test, with the conditions it held under: the author's bot, league and year. A constant from someone else's bot, an exploration weight or a velocity multiplier, is a starting point for your own tuning.
2. Build on an engine you've checked. Each bot here was measured in a simulator verified against the game's referee or recorded live games, so a local result meant something. Referees and Simulators shows how to verify one.
3. Measure against your previous bot. Use fixed maps, both seats, and enough games for an interval that excludes .5, on maps you didn't tune on. Measuring Bot Strength works out how many games that takes.
4. Test live against the boss before you submit. Play it on fixed seeds. CodinGame's own documentation adds the caution: "Beating the boss in test games is a good sign, but the ranking is computed against every player of the room".^[2]
5. Submit, record, and split the results. Save every ranked battle, then break the results down by opponent and by player count, the way Game of Drones' gap showed up.
6. Read the new league's rules after a promotion. Mad Pod Racing's thrust limit changed between Gold and Legend.

The kind of bot to start with depends on what a turn looks like. In these games it went like this:

Treat the table as a first guess for a similar game, then measure.

09Mistakes these games punished

Copying a constant. An exploration range of 0.3 to 0.5 taken from another bot never beat 0.6 in the lab's Ultimate Tic-Tac-Toe bot, and Hastings's 4 × velocity was tuned on his own pods. Tune every borrowed number on your bot.

Testing against a copy of yourself. Two deterministic bots that play the same way mirror each other: Loujo saw Game of Drones copies end 0–0 in 2014, and the lab's earlier Game of Drones bot did the same against itself in 2026.^[24][5] Measure against varied opponents. The Back Track King case study shows the same trap producing a fake win rate.

Trusting local copies of opponents. In Spring Challenge 2022 a version that beat the old one against every local replica of the shield attackers did worse than it against the real ones.^[5] Confirm a local result on a pre-registered live panel before you believe it.

Treating promotion as the end. The Mad Pod Racing bot reached Legend from the top of Gold and then ranked in the lower half of the table under the new league's rules.

Keeping a fixed opening forever. Laochao's corridor "usually won against the top players at the time", and within the same contest players learned to counter it.^[11]

10What's next

This is the last page in the series' order, so the banner goes back to the Build a Legend Bot hub. For your next game, start from the step-by-step procedure in Capstone: Build the Lab: verify an engine, write the search or rules the post-mortems point to, and measure every change before you submit it. If the game looks more like Spring Challenge 2022, with many units, hidden information and a meta to train against, the machine-learning phase starts at When to Use Machine Learning.

11Sources

Standings and measurements are the author's lab records and carry their dates. Each post-mortem is credited by its author's handle and linked to the original post; the CodinGame blog was unavailable on 2026-10-30, so its posts are linked through the Internet Archive.

1. Author's lab records, 2026 (private repository): standings. Final Gold rankings, promotions and Legend results in experiments/live-submissions/<puzzle>/rankings.jsonl and submitted.jsonl for the seven games climbed from Gold (last read October 30, 2026); Back Track King's best Legend final in evaluations-20261024.json; Fantastic Bits in AGENTS.md (October 30 entry).
2. CodinGame, 2026. Leagues and promotion, CodinGame API documentation, retrieved 2026-10-30. Bosses, promotion once your score reaches the boss's, league openings in contests, test games against the boss, and new rules after a promotion.
3. Tom Alard, 2026. Fighting for #1 in the Ultimate Tic-Tac-Toe Arena. Why naive minimax plays poorly, the ceiling of random-rollout MCTS, and networks inside the search at the top.
4. CodinGame forum, 2018 to 2026. Ultimate Tic Tac Toe [Puzzle discussion]. Magus (#33, #36) on rollouts and tree reuse, JFB (#26) on a network without search, jacek (#150, #195) on MCTS additions and the center opening, RoboStac (#162) on tree reuse.
5. Author's lab records, 2026 (private repository): what each bot does and how it was measured. docs/postmortem-lessons.md (October 29 standings and every later round); the README.md of games/tictactoe, tronbattle, madpodracing, trollfarm, spring2022 and gameofdrones; prompts/results/greatescape-legend-20261029.json and trollfarm-legend-20261029.json; experiments/trollfarm-farmer-bc-20261030.json; experiments/live-tests/<puzzle>/ (live tests against the bosses and the Spring 2022 paired panels); experiments/spring2022-attacker-20261026.json; docs/multi-game.md; experiments/runs/trollfarm/config.json.
6. Andy Sloane (a1k0n), 2010. Google AI Challenge post-mortem. Voronoi territory, counting open edges, articulation points and the chamber correction, and his view of evaluation against depth.
7. Daniel Hartmeier (dhartmei), 2010, updated 2017. Tron server with Elo ratings. Index of 2010 write-ups with final ranks: a1k0n 1st, Janzert 5th.
8. Brian Haskin (Janzert), 2010. Janzert's Tron Bots, Information Page. Alpha-beta with iterative deepening, a transposition table, bitboards and a chamber-tree evaluation.
9. Daniel Hartmeier (dhartmei), 2010. maxsteps: visiting chambers. Searching the separated endgame chamber by chamber through articulation points.
10. CodinGame forum, 2016 to 2018. Tron Battle (aka Line Racing) multiplayer challenge discussion. dolmen1234 (#36) on paranoid search over all players; Xavier_F (#34) on reaching Legend with a one-ply heuristic.
11. CodinGame forum, 2015. The Great Escape - Strategies. dimkadimon (#9) on fixed openings, Romka (#10) and Recar (#13) on their selective searches and choice of rival.
12. Aude (CodinGame), 2015. The Great Escape - Final Report, CodinGame blog (archived copy). The podium: Recar 1st, Romka 3rd.
13. CodinGame forum, 2021. Mad Pod Racing - Puzzle Discussion, post #538 by 5DN1L. Coders Strike Back is the game's previous name.
14. CodinGame forum, 2016. Coders Strike Back: Your ideas! Hastings (#22, #26), 4th: aiming at the checkpoint minus 4 × velocity, and offense or defense by race progress.
15. Magus, 2016. Coders Strikes Back, English translation by Agade. Runner and hunter roles chosen every turn from race progress.
16. pb4608, 2016. Evolutionary Trajectory Optimization with a Genetic Algorithm, CodinGame blog (archived copy). 3rd in the contest; runner and blocker roles for both players every turn.
17. delineate, 2026. Codingame Spring Challenge 2026 (delineate), GitHub gist. 1st in Troll Farm: why plain self-play never learned to mine iron, and the curriculum that fixed it.
18. Yann Moisan (yamo), 2026. My post mortem for the codingame Spring Challenge 2026. 3rd in Troll Farm: wood per turn and a two-troll economy.
19. gpoussel, 2026. Pull request #29 of the copro repository. A hand-written Troll Farm bot that rates every job by points per turn, promoted to Legend in the post-contest arena.
20. eulerscheZahl and the Troll Farm authors, 2026. Troll Farm Statistics. How many trolls each Legend player ended its games with, at the contest's end.
21. CodinGame forum, 2022. Spring Challenge 2022 - Feedbacks & Strategies. PhOeNyX (#15) on wild mana deciding games where no base falls, blasterpoard (#34), 2nd, on a bot without search, and bowwowforeach (#60, #70) on his attacker and chokudai search.
22. bowwowforeach, 2022. CodinGame Spring Challenge 2022 参加記 (in Japanese). 1st: two rule-based defenders and an attacker chosen by a 15-turn chokudai search.
23. Nanaeda, 2022. CodinGame Spring Challenge 2022で3位を取った話 (in Japanese). 3rd: all-defense until turn 120, then shielded attacks, chosen against bots that attack early.
24. Loujo, 2014. Game of Drones AI battle: Loujo's strategy (1st place), CodinGame blog (archived copy). Objectives counted in turns of travel, and the 0-0 between near-identical bots.
25. god, 2014. Game of Drones AI battle: god's strategy (2nd place), CodinGame blog (archived copy). Flexible objectives, and an edge base with three or four players.
26. Gangrene, 2014. Game of Drones AI battle: Gangrene's strategy (3rd place), CodinGame blog (archived copy). An extremity zone and its two closest neighbors.
27. CodinGame forum, 2014 to 2024. Game Of Drones multiplayer challenge discussion. tarriel (#13, 2023) on reaching Legend by rereading the post-contest threads.