Backward Induction: How Solvers Actually Think (And How You Can Too)

There is one idea sitting under every solver in the world, and almost nobody who quotes solver outputs could name it. It is not complicated. It is not new. The mathematicians figured it out long before poker software existed. And once you have held it in your hands even once, the whole machine stops looking like a black box and starts looking like a procedure you could, in a small enough spot, run yourself. The idea is called backward induction, and it is the engine that turns a decision tree into a strategy.

The problem it solves

Picture the decision tree of a hand — the whole branching structure fanning down, action after action and street after street, to the leaves where someone folds or the hand reaches showdown and the pot is awarded. If you haven't seen that picture laid out, read that piece first; here I'm going to take the tree as given and ask a single hard question about it.

The question is this: of all the choices I could make at this node, which one is best? "Best" has to mean something precise. It means the action with the highest expected value, given everything that flows forward from it. But everything that flows forward is the rest of the tree — billions of leaves. You cannot evaluate a choice at the top until you know what happens at the bottom. So where do you even start?

You start at the bottom. That is the whole trick.

Start where the values are known

At the leaves, the hand is over. The pot has been awarded. Each leaf has a value in chips that you can read off directly — no further reasoning required, because nothing else happens. That is the one place in the entire tree where the value is simply given.

So begin there. Move up one level, to the last decision made before those leaves. Ask: given both players' strategies from this point down through to the leaves, what is the expected value of each possible action at this node? The expected value is just the weighted average of all the outcomes that flow forward from that action, weighted by their probabilities under both strategies. And — this is the part to sit with — it is fully computable, because everything below this node is already known. The leaves were known directly. So the level just above them is now known too.

Pick the action with the highest expected value at that node. That is the optimal action there, given everything that follows.

Climb level by level

Now move up another level. You already know the optimal action at every node below you, so you can compute the expected value of each option at this earlier node the same way — weighted average over the outcomes that flow forward, except now those outcomes already have their values pinned down by the work you did one level lower. Pick the best one again.

Keep going. Up another level, and another. Each step is the same step: look at the layer below, which you have already solved, take the weighted average for each action, choose the highest. The tree gets solved from the floor up, one storey at a time. By the time you reach the root, you have constructed a complete strategy that maximizes expected value across the whole game — built entirely by working backward from the leaves through every node in the structure.

That procedure is backward induction. And at its mathematical core, it is exactly how every solver in poker works under the hood. The solver is not doing anything mystical. It is doing this, at a scale and speed your nervous system cannot match, over a tree too big for any human to hold in working memory.

Why you cannot run it at the table

Here is the honest limit. The full tree is too big. The leaves number in the billions. The probabilities are too complex, and they branch at every chance node and every information set. You are never going to run complete backward induction in your head in real time, and any course that pretends otherwise is selling you something.

So the practical question — the one that fills the rest of the real work — is what shortcuts, what intuitions, what structural understandings let you act as if you had run the backward induction, even though you obviously cannot. The shortcuts are real. They have names and shapes and textures: polarization, bluff-to-value ratios, minimum defense frequencies, the indifference principles. Every one of them is, at its root, a way to approximate in real time the answer that backward induction would have given you. They are compressions of the tree. And the player who has internalized them is, in effect, running a compressed version of backward induction in his body — delivering an approximate answer without the calculation ever becoming conscious.

Where you actually can run it: the river

There is one place where the tree is small enough that you can do the backward induction explicitly, with the naked eye, in something like a usable amount of time. The river.

Imagine a river situation. The last card has fallen, and now there are only two decisions left in the whole game. You go first. You can check or bet. If you check, the hand goes to showdown and one of you wins based on your hands. If you bet, your opponent can call or fold. If she folds, you win the current pot. If she calls, it goes to showdown for the larger pot. That is the entire remaining tree — a couple of decisions, a handful of leaves. The whole rest of the hand collapsed into a small finite structure.

And the math on that small structure is fully tractable. What is the value of betting this specific hand? With some probability she folds, and you win the current pot. With some probability she calls, and then you either win or lose the showdown depending on whose hand is better. Multiply, add, weighted average — there is the expected value of betting. Now compare it to the expected value of checking, which is just the showdown value of your hand against her checking range. Whichever is bigger is the right action.

That is it. You have just done backward induction on a small part of the tree, in your head. You started at the leaves — fold, you win the pot; call-and-win, call-and-lose, showdown values — and you climbed one level up to the bet/check decision and chose the higher-EV branch. The same procedure the solver runs across billions of leaves, run by you across four or five.

Why doing it once matters

You could just look up the answer. The solver will tell you to bet or check, and it will be right. But the looking-up and the deriving are not the same act, and they do not leave the same thing behind in you. When you do the backward induction yourself, even with rough numbers, you install a piece of intuition that no solver output can give you. You start to feel why the bet is better, not just that it is. And that feel is what generalizes — to the turn, the flop, the preflop, where the same structure operates in more complex form and where you will never have a clean answer handed to you.

The solver can give you answers. It cannot give you the intuition. The intuition is grown one hand at a time, by sitting with the structure on the small tractable parts of the tree — the rivers, the simple end games — until it starts to dwell inside you. Backward induction is the name of that structure. The river is where you can finally see it work.

---

This essay is drawn from the audio lesson Break Through to the Next Dimension — hear the whole argument.