Hiss NN · Documentation
The Network
A 259,599-parameter net that turns 86 numbers about a poker situation into an
action, a bet size and a value estimate, in about a millisecond on a CPU. This documents what it
is, every idea it rests on, and how to rebuild the whole thing from an empty machine.
CFR, regret, abstraction, Deep CFR, PPO — every idea the system uses
The milestone ladder, each with its gate, in rebuild order
The served net: 86 features, three heads, and the exact contract
From an empty machine to a serving champion, step by step
What it is, in one paragraph
Hiss plays no-limit Texas hold'em by reading a table off a phone screen, turning what it sees
into a fixed-length vector of 86 numbers, and asking a neural network what to do. The network is a
blueprint: it has already been trained to approximate good play and does no
searching at decision time, because the decision budget is milliseconds on a CPU that is
simultaneously running computer vision for several handsets. Everything in this documentation
exists to answer one question — how do you get a net whose blueprint is worth trusting?
The two tracks, and why there are two
The system has been built along two independent paths that meet at the same serving contract.
Understanding that they are different is the single most useful orientation.
| Track A — imitation & RL | Track B — game theory |
|---|
| Idea | Copy good play, then improve it by trial | Approximate an equilibrium from first principles |
| Methods | Behavioural cloning → PPO | CFR → Deep CFR → distillation |
| Needs | A teacher or a reward signal | A simulator of the rules |
| Fails by | Inheriting the teacher's mistakes; over-fitting the opponents it trained against | Abstraction error; cost; being unexploitable but not maximally profitable |
| Milestones | The live AIL/PPO cycle | M0 → M6 |
Why both
An equilibrium policy is hard to beat but does not exploit: it wins little from bad
players because it is not trying to. An imitation/RL policy exploits well but drifts toward
whatever it was trained against — and self-play opponents that never fold will teach a net to be a
maniac. The practical answer is to use the game-theoretic track for the shape of
the policy and the RL track for adaptation, with a gate that only promotes on
measured money against realistic opponents.
How one decision is made
End to end, live, on every hand:
- See. A capture chain publishes a frame of the phone screen. Computer vision
(OpenCV plus OCR) reads cards, stacks, bets, seats, buttons.
- Featurise. The table state becomes 86 floats — board texture, made-hand and
draw flags, position, stack-to-pot ratio, pot odds, opponent statistics, hole-card shape, equity.
The full list is on the architecture page.
- Normalise. Each feature is standardised with the
mu and
sd vectors saved beside the weights, so the net sees the same distribution it was
trained on.
- Infer. A shared tower produces a 192-dimensional representation; three heads
read it — policy (6 action classes), bet size (8 pot fractions),
value (a scalar estimate).
- Constrain. Serve-time guardrails can override the net where it is known to be
weak — a preflop opening rail, a fold-to-three-bet discipline, a short-stack push/fold policy.
- Act. The chosen action becomes a tap on the glass, with a read-back check that
the tap actually registered.
The guardrails are load-bearing, and they distort measurement
Guardrails exist because a net can be excellent in general and reliably wrong in a specific
spot. But a guardrail that overwrites the net also hides it: when a preflop rail overrode the net on
roughly three quarters of preflop decisions, most challengers scored exactly +0.00 against
the champion, because the two models were never allowed to differ. A gate measuring a policy that
never plays is measuring nothing. Always report how many decisions a treatment actually changed.
What "good" means here
Three different measurements, and they answer different questions:
| Measure | Question it answers | Limit |
|---|
| Exploitability | How much could a perfect opponent win from me? | Only computable exactly in small games |
| LBR | How much can a simple, greedy opponent win? | A lower bound — safe to trust in one direction only |
| bb/100 | How much money does it actually win, per hundred hands? | Enormously noisy; needs tens of thousands of hands |
The last one is the arbiter, because it is the only one denominated in the thing that matters —
but it is also the one that most readily lies. Variance in poker is large enough that a losing
policy can outperform a winning one over thousands of hands, which is why every promotion decision
here is a confidence interval rather than a comparison of two numbers.
Next — every concept the system uses →
