Honest Neural Tic-Tac-Toe V5

Train

idle

Custom games Chunk

Games

0

—

Temperature

1.10

state-aware exploration

Learning rate

0.0015

auto stability tuning

Eval score

—

vs random, no oracle

Recent W/D/L

0/0/0

training games

Policy entropy

—

exploration health

Experience rows

0

state/action memory

Canonical states

0

after symmetry folding

Self-play

40%

Learn as X vs random

20%

Learn as O vs random

20%

Extra exploration

20%

Purity guard: legal masking and symmetry folding are allowed because they only encode the game representation. The trainer never asks what the best move is and never rewards “blocking”, “taking center”, or any other tactic. Random-opponent moves are no longer recorded as learner targets; only the learner’s own choices update its experience table.

Play / inspect

your turn

You play AI brain AI play temp 0.03 Show AI suggestions

Train a little, then challenge the network. The bars inside empty cells are the model's current move probabilities.

Speed 260 ms

Training log

What this learner is allowed to know

Honest training contract

The model sees the board from the current player's perspective, chooses a legal move, then receives only the final outcome: +1 for win, 0 for draw, -1 for loss.

Adaptive reward is implemented as final_result - predicted_value. That means a draw can be good or bad depending on what the model expected, without any tactical hint being inserted.

Board positions are folded through the 8 tic-tac-toe symmetries. This reduces duplicate learning but does not tell the network how to play.

V5 change: the learner is unchanged from V4, but the game board is now a locked 3×3 square grid, AI move suggestions can be hidden during play, and the current model can be watched playing both sides against itself.

V4 learner: training is deliberately balanced between playing first and second. The learner explores under-tested actions early, but once a move has repeatedly performed badly in that exact canonical position, its exploration probability is reduced rather than endlessly repeated.