Hiss NN · 03
Architecture
The contract. Any training path — imitation, PPO, CFR distillation — is only
useful if it produces something in exactly this shape, because this is what the serving process,
the bots and the drivers already expect.
The network
input 86 features (standardised: x' = (x - mu) / sd, from norm.npz)
|
tower Linear(86 -> 384) + LayerNorm(384) + activation
Linear(384 -> 384) + LayerNorm(384) + activation
Linear(384 -> 192)
|
+-- policy Linear(192 -> 6) action class logits
+-- betsize Linear(192 -> 8) pot-fraction bucket logits
+-- value Linear(192 -> 1) scalar state value
total parameters: 259,599
Deliberately small. It has to run on a CPU that is simultaneously doing computer vision for
several handsets, inside a decision budget measured in milliseconds, so capacity is spent on the
feature engineering rather than on depth.
Why three heads on one tower
The three questions — what to do, how much, and how good is this spot — share almost all of their
evidence. One shared representation means the value head's learning signal also improves the policy
head's features, and the value head is what PPO needs to compute advantage. Separate networks would
triple the cost and learn the same thing three times.
Action space
| Head | Width | Classes |
|---|
| policy | 6 | fold · check · call · raise · allin · bet |
| betsize | 8 | 0.33 · 0.5 · 0.75 · 1.0 · 1.5 · 2.5 · 3.0 · 4.0 × pot |
| value | 1 | scalar |
The bet head was trained on a size that never happened
For a long period the bet head learned from data in which a serve-time guardrail overwrote every
bet the network chose. It was therefore fitting a size it had never actually played, and no metric
said so. It only came right when the action space, the objective and the optimiser were fixed
together; fixing any one alone did nothing measurable.
Note also that 3.0 and 4.0 were appended to the bucket
list rather than inserted, so older six-wide champions still decode correctly. Widening an ordinal
head in the middle silently relabels every historical model.
The 86 features
Grouped by what they describe. The order is the contract — norm.npz's
mu and sd are positional, so inserting a feature anywhere but the end
invalidates every existing checkpoint.
Board texture — 0…5
00 f$BoardWet
01 f$BoardDry
02 f$ScaryBoard
03 f$BoardHighCardFoldy
04 f$BoardHighCardSticky
05 f$BoardParched
Made hands & draws — 6…19
06 HaveTopPair
07 HaveOverPair
08 HaveSet
09 HaveTwoPair
10 HaveFlush
11 HaveStraight
12 FlushPossible
13 StraightPossible
14 f$HaveStrongMade
15 f$HaveBigMade
16 f$HaveOnePair
17 f$HaveComboDraw
18 f$HaveBigDraw
19 f$HaveWeakDraw
Position & stack — 20…31
20 f$InPositionPost
21 f$InPositionPre
22 f$SPR
23 f$Committed
24 pot_odds
25 f$M
26 f$Mzone
27 f$DeepStack
28 f$ShortStack
29 f$PushFoldStack
30 f$HeadsUpPot
31 f$Is6Max
Action state — 32…39
32 Raises
33 Calls
34 Bets
35 betround
36 iamaggressor
37 isplo8
38 isomaha
39 prwin
Opponent model — 40…56
40 f$Opp_VPIP
41 f$Opp_PFR
42 f$Opp_AF
43 f$Opp_WTSD
44 f$Opp_Hands
45 f$Opp_Known
46 f$Opp_IsNit
47 f$Opp_IsLoose
48 f$Opp_IsStation
49 f$Opp_IsPassive
50 f$Opp_IsAggro
51 f$Opp_IsTAG
52 f$Opp_IsLAG
53 f$Opp_IsFish
54 f$Opp_IsManiac
55 f$Opp_Foldy
56 f$Opp_ThreeBetsLight
Hole-card shape — 57…62
57 hole_hi
58 hole_lo
59 hole_suited
60 hole_pair
61 hole_gap
62 hole_broadway
Exploit reads — 63…68
63 f$Opp_CBet
64 f$Opp_FTC
65 f$Opp_F3B
66 f$Opp_Steal
67 f$Opp_FTS
68 f$Opp_ExploitKnown
Seat & steal — 69…76
69 f$FoldedToMe
70 f$IsBTN
71 f$IsCO
72 f$IsSB
73 f$IsBB
74 f$IsEP
75 f$PlayersLeftToAct
76 f$StealSpot
Players behind — 77…80
77 f$Behind_FTS
78 f$Behind_VPIP
79 f$Behind_ThreeB
80 f$Behind_Known
Range equity — 81…82
81 prwin_range
82 prwin_gap
Street memory — 83…85
83 f$ICheckedStreet
84 f$IBetStreet
85 f$CheckRaiseSpot
How to read this list
The f$ prefix marks features that come from the rule-engine's own vocabulary — the
same names its book uses — which is what lets a rule interpreter produce training labels that line
up with the network's inputs. The rest are computed directly from the table read. Note how much of
the vector is opponent model (17 features) and seat/steal context
(8): the hand itself is a minority of what the network sees, which is correct for poker.
Feature gates: the marker files
Feature groups are switched on by the presence of a marker file beside the weights, so a champion
records which capabilities it was actually trained with:
| Marker | Enables |
|---|
SIGHTED | multiway equity — the net can see it is not heads-up |
AGG_TRAINED | iamaggressor — who took the initiative |
TRAP_TRAINED | street memory — what I did earlier this street |
EXPLOIT_TRAINED | exploit / steal / players-behind reads |
A feature the policy cannot see is not a feature
Exploit features were built, deployed and measured — and were invisible to the policy:
adding them changed nothing, because the learning path never credited them. Adaptation itself was
worth +196 bb/100 when it worked, so the failure was in learning, not in the features. A
marker file records that a group was present; it does not prove the net used it.
Serving
| Service | Port | Role |
|---|
hiss-nn-decide | 8088 | Champion — the live net |
hiss-nn-decide-chal | 8091 | Challenger — the A/B candidate |
hiss-nn-decide-cfr | — | CFR net, its own A/B slot |
hiss-decide | 8087 | Rule engine, headless — the teacher and baseline |
Artefacts live together and travel together:
model_champion/
model.pt raw weights
model_scripted.pt TorchScript — what is actually served
norm.npz mu, sd (86 each) — positional, order is the contract
profile.json the style this champion plays (vpip, pfr, af, wtsd, 3bet, …)
SIGHTED AGG_TRAINED TRAP_TRAINED EXPLOIT_TRAINED feature gates
Guardrails
Rules that override the network at serve time, each added because the net was reliably wrong in
one specific place and could not be trained out of it:
- Preflop opening rail — constrains opening frequency and size.
- Fold-to-three-bet discipline — worth +13.3 bb/100 as a guardrail. It could not
be trained: the net would not hold the behaviour.
- Short-stack push/fold — the net was trained only at ~100bb and limped aces at
8bb, which is a training-distribution hole, not a policy error.
- Anti-donk — a champion donk-shoved because it was postflop-blind on a sparse
symbol path and saw zeros.
Guardrails blind the gate that judges the net
The preflop rail overwrote the net on ~75% of preflop spots, so 87 of 114 PPO
challengers scored exactly +0.00 against the champion. The gate was not being strict; it
was being blind. Any guardrail that improves play also removes the evidence that would let
you improve the policy underneath — so measure how often each one fires.
Next — rebuild it from an empty machine →
