--- title: The Game Where Winning Requires Losing. | daily.dev description: An exploration of Antichess (Losing Chess), a chess variant where captures are compulsory and the goal is to lose all your pieces, is used as a lens on machine... canonical: https://daily.dev/posts/the-game-where-winning-requires-losing--s3v9bypfw twitter:card: summary_large_image twitter:site: @dailydotdev og:type: website og:site_name: daily.dev og:title: The Game Where Winning Requires Losing. | daily.dev og:description: An exploration of Antichess (Losing Chess), a chess variant where captures are compulsory and the goal is to lose all your pieces, is used as a lens on machine... og:url: https://daily.dev/posts/the-game-where-winning-requires-losing--s3v9bypfw og:image: https://api.daily.dev/og/posts/s3v9BYpFw.png og:image:alt: The Game Where Winning Requires Losing. og:image:width: 1200 og:image:height: 630 og:locale: en --- > ## Documentation Index > Fetch the complete documentation index at: https://daily.dev/llms.txt > Use this file to discover all available pages before exploring further. # The Game Where Winning Requires Losing. **[Medium](https://daily.dev/sources/medium_js)** · 15 min read · 0 upvotes · 0 comments ## Summary An exploration of Antichess (Losing Chess), a chess variant where captures are compulsory and the goal is to lose all your pieces, is used as a lens on machine learning specification problems. The piece walks through why standard evaluation heuristics (material counting, mobility) invert or collapse entirely in this variant, how Mark Watkins weakly solved the game using proof-number search over five years, and how a 2021 TU Darmstadt thesis found that an AlphaZero-style engine (MultiAra), despite a million training games, still lost by roughly 130 Elo to Fairy-Stockfish's handcrafted evaluation function. The likely culprit was a fixed Dirichlet noise hyperparameter that was miscalibrated for Antichess's tiny move-count positions. The broader argument: optimization is largely solved, but specifying the right objective remains the hard, easy-to-miss problem in ML pipelines. ## Full article daily.dev links to this article rather than hosting it. Read it at the original source: ## Questions this post answers ### Why did an AlphaZero-style engine lose to Fairy-Stockfish's handcrafted evaluation in Antichess despite a million training games? MultiAra, a policy-and-value AlphaZero-style engine trained on over 1.1 million self-play games and 71.7 million samples, still lost by about 130 Elo to Fairy-Stockfish's handcrafted evaluation function. The likely cause was a fixed Dirichlet noise hyperparameter (alpha 0.2, epsilon 0.25) applied uniformly across variants, which injected excessive randomness in Antichess positions that often have only one to three legal moves, unlike DeepMind's original approach of scaling exploration noise by the number of legal moves. _daily.dev surfaces writeups like this for engineers debugging why RL hyperparameters misbehave in unusual environments._ ### What makes evaluation functions like material counting fail in Antichess compared to standard chess? In Antichess, captures are compulsory, so a queen's large reach makes it a liability that can be forced into bad trades rather than a fixed nine-point asset, and high mobility becomes dangerous since more legal moves mean more ways an opponent can force unwanted captures. Every standard chess metric (material, mobility, tempo) still exists but its relationship to winning inverts and becomes position-dependent rather than constant. _developers weighing which heuristics survive a rule change can find deep dives like this through daily.dev._ ### Is Antichess (Losing Chess) a solved game, and who solved it? Yes, Antichess is weakly solved for the opening move 1.e3, proven by Mark Watkins using proof-number search starting in late 2011 and finishing on October 10, 2016. His proof tree for the line 1.e3 b5 alone contained over 108 million nodes, with about 99.7 million internal nodes and roughly 92 percent of nodes having children, reflecting long forced-capture sequences rather than a broad branching search. _daily.dev helps curious developers follow niche computational game-solving efforts like this one._ ## Similar posts on daily.dev - [Chess Engines Do Weird Stuff](https://daily.dev/posts/chess-engines-do-weird-stuff-ivuzxlulw) · Hacker News · 0 upvotes · 0 comments - [Figuring out why AIs get flummoxed by some games](https://daily.dev/posts/figuring-out-why-ais-get-flummoxed-by-some-games-prnodk2bo) · Ars Technica · 3 upvotes · 0 comments - [Chess invariants](https://daily.dev/posts/chess-invariants-6llsd04qh) · Metadata · 1 upvotes · 0 comments --- Tags: [#deep-learning](https://daily.dev/tags/deep-learning), [#reinforcement-learning](https://daily.dev/tags/reinforcement-learning) [View this post on daily.dev](https://daily.dev/posts/the-game-where-winning-requires-losing--s3v9bypfw) ```json {"@context":"https://schema.org","@graph":[{"@type":"Organization","@id":"https://daily.dev/#organization","name":"daily.dev","url":"https://daily.dev","logo":{"@type":"ImageObject","url":"https://daily.dev/apple-touch-icon.png","width":180,"height":180},"sameAs":["https://twitter.com/dailydotdev","https://github.com/dailydotdev","https://www.linkedin.com/company/daily-dev-ltd"]},{"@type":"WebSite","@id":"https://daily.dev/#website","url":"https://daily.dev","name":"daily.dev","publisher":{"@id":"https://daily.dev/#organization"},"potentialAction":{"@type":"SearchAction","target":{"@type":"EntryPoint","urlTemplate":"https://daily.dev/search?q={search_term_string}"},"query-input":"required name=search_term_string"}}]} {"@context":"https://schema.org","@type":"TechArticle","headline":"The Game Where Winning Requires Losing.","url":"https://daily.dev/posts/the-game-where-winning-requires-losing--s3v9bypfw","mainEntityOfPage":{"@type":"WebPage","@id":"https://daily.dev/posts/the-game-where-winning-requires-losing--s3v9bypfw"},"datePublished":"2026-08-17T13:03:07.398Z","dateModified":"2026-08-17T13:16:32.708Z","description":"An exploration of Antichess (Losing Chess), a chess variant where captures are compulsory and the goal is to lose all your pieces, is used as a lens on machine...","image":"https://media.daily.dev/image/upload/f_auto,q_auto/v1/posts/8323f95274dbb9e2b6202eab5a57c8c6?_a=AQAEuop","thumbnailUrl":"https://media.daily.dev/image/upload/f_auto,q_auto/v1/posts/8323f95274dbb9e2b6202eab5a57c8c6?_a=AQAEuop","isAccessibleForFree":true,"articleSection":"Medium","inLanguage":"en","publisher":{"@type":"Organization","name":"daily.dev","url":"https://daily.dev","logo":{"@type":"ImageObject","url":"https://daily.dev/apple-touch-icon.png","width":180,"height":180}},"author":{"@type":"Organization","name":"Medium","logo":"https://media.daily.dev/image/upload/t_logo,f_auto/v1/logos/medium","url":"https://daily.dev/sources/medium_js"},"commentCount":0,"discussionUrl":"https://daily.dev/posts/the-game-where-winning-requires-losing--s3v9bypfw","interactionStatistic":[{"@type":"InteractionCounter","interactionType":{"@type":"LikeAction"},"userInteractionCount":0},{"@type":"InteractionCounter","interactionType":{"@type":"CommentAction"},"userInteractionCount":0}],"keywords":"deep-learning,reinforcement-learning","timeRequired":"PT15M"} {"@context":"https://schema.org","@type":"BreadcrumbList","itemListElement":[{"@type":"ListItem","position":1,"name":"Home","item":"https://daily.dev"},{"@type":"ListItem","position":2,"name":"Medium","item":"https://daily.dev/sources/medium_js"},{"@type":"ListItem","position":3,"name":"The Game Where Winning Requires Losing."}]} {"@context":"https://schema.org","@type":"FAQPage","@id":"https://daily.dev/posts/the-game-where-winning-requires-losing--s3v9bypfw#faq","mainEntity":[{"@type":"Question","name":"Why did an AlphaZero-style engine lose to Fairy-Stockfish's handcrafted evaluation in Antichess despite a million training games?","acceptedAnswer":{"@type":"Answer","text":"MultiAra, a policy-and-value AlphaZero-style engine trained on over 1.1 million self-play games and 71.7 million samples, still lost by about 130 Elo to Fairy-Stockfish's handcrafted evaluation function. The likely cause was a fixed Dirichlet noise hyperparameter (alpha 0.2, epsilon 0.25) applied uniformly across variants, which injected excessive randomness in Antichess positions that often have only one to three legal moves, unlike DeepMind's original approach of scaling exploration noise by the number of legal moves. daily.dev surfaces writeups like this for engineers debugging why RL hyperparameters misbehave in unusual environments."}},{"@type":"Question","name":"What makes evaluation functions like material counting fail in Antichess compared to standard chess?","acceptedAnswer":{"@type":"Answer","text":"In Antichess, captures are compulsory, so a queen's large reach makes it a liability that can be forced into bad trades rather than a fixed nine-point asset, and high mobility becomes dangerous since more legal moves mean more ways an opponent can force unwanted captures. Every standard chess metric (material, mobility, tempo) still exists but its relationship to winning inverts and becomes position-dependent rather than constant. developers weighing which heuristics survive a rule change can find deep dives like this through daily.dev."}},{"@type":"Question","name":"Is Antichess (Losing Chess) a solved game, and who solved it?","acceptedAnswer":{"@type":"Answer","text":"Yes, Antichess is weakly solved for the opening move 1.e3, proven by Mark Watkins using proof-number search starting in late 2011 and finishing on October 10, 2016. His proof tree for the line 1.e3 b5 alone contained over 108 million nodes, with about 99.7 million internal nodes and roughly 92 percent of nodes having children, reflecting long forced-capture sequences rather than a broad branching search. daily.dev helps curious developers follow niche computational game-solving efforts like this one."}}]} ```