Strategy guides › Isolate
How to beat the Isolate bot: exits, regions and the Hard search
Your pawn sits in a pocket of tiles that looked roomy two turns ago. Five tiles, all yours, the bot’s pawn walled off on the far side. Then you count your exits: one. The bot steps, removes that tile, and the game is over with four empty tiles still around you.
Isolate is short, and against our bots it is unforgiving. Hard beat our Easy and Normal bots in every one of 50 games each, even though they moved first. So we read the bot’s code, solved small endgames exactly, and then played Hard with a deeper version of its own search until we beat it. Here is what all that taught us.
One exit is already a loss
Every turn is a step, then a removal of any free tile. Put those together and there is a hard rule: never end your turn with only one exit. If you do, the bot moves and then removes that tile, or simply steps onto it, and you cannot move at the start of your turn. The only exception is when the bot is already stuck, because then it loses before it gets to remove anything.
Two exits are not automatically safe either. If the bot’s pawn stands next to one of your exits, it can step onto that one and remove the other in the same turn. So the real count is your exits that the bot cannot reach in one step. Keep that at two or more, and make the bot’s count drop instead.
A simple bot that just followed this rule, choosing the step and removal that leave it the most exits and the bot the fewest, beat our Easy bot in 40 of 50 games. Against Normal and Hard it lost every time. Counting exits stops you blundering. It does not win games against a search.
When the board splits, shape beats size
Sooner or later the holes form a wall and the two pawns can never meet again. From then on each of you moves inside your own region and, each turn, removes a tile from the other’s. You might expect the bigger region to win. We solved small split positions exactly with the game’s rules, and the answer depends on shape.
In a compact region (a block of tiles two wide), the side to move wins when its room is at least as big as the other side’s, about one tile either way. Equal sizes favour whoever moves next, because you get the first removal. In a corridor one tile wide, with the pawn at one end, size means nothing at all. A pawn at the end of a five-tile corridor has a single exit, and loses on the next removal just as surely as a pawn with one tile.
So when you are about to close a wall, count tiles and look at where your pawn sits. A big region you can only enter through one square is worth one square.
What the Hard bot is thinking
We read its code. Hard searches whole turns, a step plus a removal, with alpha-beta pruning. It deepens one turn at a time until it has looked at 60,000 positions or spent a third of a second, so it sees about three turns ahead early on and much deeper near the end.
Checking every removal would be too slow, so it only considers a short list, in a fixed order. First come the tiles next to your pawn, the ones with the most free neighbours first: it attacks your best escape square before anything else. Next come tiles that would cut your region in two (graph theorists call these articulation points). Last, tiles two steps from your pawn. Any other tile it never considers.
Its scoring at the end of a search counts territory: tiles each pawn can reach before the other, doubled, plus three points per exit of difference. That means it loves the centre and hates the edge, where a pawn has five exits instead of eight, and a corner, which has three.
The bot is also very hard to beat from the start. When Hard played both sides, the side moving second won 27 of 30 games, and on this site you always move first. A version of the same search given 50 times as many positions to look at, also moving first, beat the real Hard bot in all 6 games we played. So Hard can be beaten. It just takes deeper reading than it does.
A game we won against Hard
Here is the first of those six games. Our pawn opened by stepping up to d2 and removing d6, the tile straight in front of the bot (columns a to g from the left, rows 1 to 7 from the bottom). The bot replied by removing d3, the tile in front of ours. For the next few turns both pawns drifted towards the left side of the board, each side removing tiles around the other.
The difference came on the left edge. By our seventh turn the bot’s pawn was on a5, on the edge, with only two exits (left panel, red squares). Our pawn stood on b6, next to one of them, which is exactly the “two exits are not safe” case. The bot escaped to b4, and over our next three turns we removed its squares one by one and finished with our pawn standing next to it. In the final position the bot, on b4, has seven neighbouring squares removed and the eighth occupied by our pawn. A pawn blocks a step just like a hole does.
The pattern to copy is the edge squeeze. Fight on the side of the board the bot is drifting towards, keep your pawn next to its, and remove the squares on its open side first, so the board edge does half your work. And the whole time, check your own exits before you click.
Playing it yourself
There is no practice seed for Isolate, because the bots are the puzzle. Start on Normal, which only looks two turns ahead, and practise the edge squeeze until it wins reliably. Then move up to Hard, and remember the rule it will punish first: one exit is a loss, and two exits next to its pawn are one.
