Monopoly vs Computer: How the AI Bots Play and How to Beat Them

You can play Monopoly against the computer here for free, in every seat you leave empty, and the opponent you get learned the game from 80,000 full matches instead of a list of rules. Easy loses to a careful beginner. Normal is a fair fight. Hard beat our previous strongest bot in 83 of every 100 games we ran between them.

How the bot decides a move

Before every decision the bot lists each legal move it has, plays each one out on a private copy of the board, and asks a small neural network one question about the result: from this seat, how good is this position? Then it plays the move that scored highest. That's the whole method, and it covers buying, bidding, building, mortgaging, trading and the casino with the same question every time.

The network's sense of "good" comes from 80,000 games between our older, hand-written bots, 5.8 million positions in all, each one labelled with how the game eventually ended for the player sitting there. Nobody told it which colour groups pay. It worked that out from which seats kept finishing first. Gerald Tesauro built the backgammon program TD-Gammon on the same idea in the early 1990s (Communications of the ACM, March 1995): rate positions, don't write rules, and let a search over the next move do the rest.

The limit is worth knowing too. It looks one move ahead, not three turns ahead, which is why it's sharpest in a duel and softer at a full table.

What easy, normal and hard change

One network drives all three levels. The difficulty setting changes how clearly the bot reads it.

An easy bot sees the board through fog: its read of every cash balance and every rent threat is nudged by about ten percent at random, it picks loosely among moves that score close together, and one move in twenty is random outright. It never proposes a trade, stops building at two houses, and will take a deal that's slightly against it.

A normal bot reads the board almost cleanly and picks the best move most of the time. It offers cash for a street that completes its set and nothing fancier, builds to three houses, and never overpays at auction to block you.

A hard bot plays the top-rated move nearly every time. It builds to hotels, constructs swaps and two-for-one offers, bids past a street's worth purely to keep you off a set, and accepts a trade only when its position comes out ahead.

We measured each level in 1,200 duels against the bot that used to be our hardest setting. Easy won 26.2% of them, normal 52.8%, hard 81.8%. Normal beats easy 74 times in 100, and hard beats normal 74 times in 100. For scale, the people playing here beat that older bot in 61% of their played-out duels over the last 90 days, so easy is comfortably beatable, normal is a fair fight, and hard is meant to be hard.

Across 5,704 bot seats filled so far, players picked normal 4,413 times, easy 661 and hard 630.

How it buys, bids and builds

With 8,000 or more left after the purchase it simply buys, the way the older bot always did in the opening. Below that line it says no to about half the streets it lands on: in 200 test duels against the older bot, the new hard bot declined 2,064 of 4,273 purchase offers, which looked like a bug until we counted the older bot doing the same, 1,751 of 3,938. Seven in ten of the streets it declined went unsold at auction, and when the older bot did take one it paid 1.05 times face for it. A decline with cash that tight is a judgement that the street isn't worth face to it right now, not a lapse.

Building is where the two bots part ways. The new one built 4,691 times across those 200 games against the older bot's 1,657, about 23 houses a game against 8, and it starts earlier. Houses are where the rent is, and the network rates a set with three houses far above the same set bare.

At auction it raises one step of 100 at a time and stops the moment "own this at that price" scores below "let it go". Hard adds the blocking bid on top: if a street would complete your group, it keeps bidding past what the street is worth to it. Our auction guide covers what that does to your side of the table.

Whether it will trade with you

All three levels answer your proposals. Normal and hard also open trades of their own, normal with cash offers only and hard with anything up to a two-for-one. Easy never proposes.

The new bot asks less and agrees more. In the same 200 duels it proposed 327 trades to the older bot's 707, about three offers every two games, and it accepted 247 of the 707 proposals it received, one in three, where the older bot said yes to 75 of 327, fewer than one in four. A proposal is accepted when the board after the trade scores higher for the bot than the board before it, with a small margin on hard. Hand a bot a deal that completes your monopoly and the after-board scores badly for it, however much cash you attach, unless what comes back is worth more to it than what you gain. The trading guide shows how to shape an offer that survives that test.

How to beat each level

Easy never blocks an auction. Decline the streets you don't want and pick up the ones you do when they come around at auction, then build to three houses on one group. It stops at two, and rent at three houses is where its cash runs out.

Normal doesn't block either, and it only ever offers cash. Its blind spot is time: it rates the board right after the trade, not the board two turns later, so a swap that's dead even on paper and lets you build immediately is a swap it will accept. Offer even trades that hand you tempo.

Hard is the one that costs you money to bait, and the bait works. Bid up any street that would complete your set, because hard will pay over the odds to keep it from you, and every 100 it spends blocking is 100 it can't put into houses. Hard is also at its strongest one on one. With three opponents its one-move lookahead is worth less, which showed in our tests: the same bot that wins 83% of duels wins 32.5% of four-player tables, better than the 22.4% an older bot manages at the same table but a long way from unbeatable.

One more thing you'll notice: the bot bets at the casino when it has cash to spare. The payout table there returns the stake on average at every bet size and adds a jackpot draw, so betting is a question of whether you can afford the swing, and the bot answers it the same way you should.

Team games keep the older bot

In 2v2 team games the learned bot did not beat the old one. It won 46.9% of 1,500 team games, so team tables still run the hand-written bot, and we'll switch them over only when the new one is better there. The team rules are unchanged.

What a bot game does and doesn't give you

No rating, no experience, no recorded win. A game with any bot in it is casual by design, because a ladder that can be farmed against a machine stops being a ladder. The game still lands in your history and your daily streak still counts.

Bots never play to an empty room. Close the tab and the table freezes where it stands, and an abandoned bot table is cleared after 30 minutes.

How long a game against the computer takes

Thirty minutes, at the median. Over the last 90 days, 974 finished bot games that reached at least ten rounds took a median of 30 minutes, against 39 minutes for the 1,611 human-only games of the same length, and both ran about 58 rounds. The bot pauses between one and four seconds before each move so you can see the dice; the rating itself takes a few thousandths of a second. Last week 346 bot games finished on this site. Our game length breakdown has the human numbers in full.

A hard bot wins 83 games in 100 against the opponent most players here were already losing 4 in 10 to. Pick a level in the lobby and find out which one you are.