Spaces:
Running
Running
File size: 7,145 Bytes
502af73 |
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 |
/**
* Trigo AI Agent - Language Model-based Move Selection (Frontend/Backend Common)
*
* Platform-agnostic AI agent that accepts ONNX session from platform-specific code.
* No direct dependency on onnxruntime packages - uses dependency injection pattern.
*
* Uses ONNX language model to score and select moves by:
* 1. Getting all valid moves for current position
* 2. Scoring each move by appending it to TGN and computing token probabilities
* 3. Selecting move with highest probability (argmax)
*/
import { ModelInferencer } from "./modelInferencer";
import { TrigoGame, StoneType } from "./trigo/game";
import type { Move, Stone } from "./trigo/types";
/**
* Configuration for the AI agent
*/
export interface TrigoAgentConfig {
vocabSize?: number;
seqLen?: number;
temperature?: number;
}
/**
* Move score result
*/
export interface MoveScore {
move: Move;
score: number;
logProb: number;
}
/**
* Trigo AI Agent for move generation
* Compatible with both frontend (onnxruntime-web) and backend (onnxruntime-node)
*/
export class TrigoAgent {
private inferencer: ModelInferencer;
constructor(inferencer: ModelInferencer) {
this.inferencer = inferencer;
}
/**
* Check if agent is initialized (checks if inferencer has a session)
*/
isInitialized(): boolean {
// Agent is initialized if the inferencer has been set up
return this.inferencer !== null;
}
/**
* Convert Stone type to player string
*/
private stoneToPlayer(stone: Stone): "black" | "white" {
if (stone === StoneType.BLACK) return "black";
if (stone === StoneType.WHITE) return "white";
throw new Error(`Invalid stone type: ${stone}`);
}
/**
* Convert string to token IDs (byte-level encoding)
*/
private stringToTokens(text: string): number[] {
return Array.from(text).map((char) => char.charCodeAt(0));
}
/**
* Compute softmax probabilities from logits
*/
private softmax(logits: Float32Array, vocabSize: number): Float32Array {
const probs = new Float32Array(vocabSize);
let maxLogit = -Infinity;
// Find max for numerical stability
for (let i = 0; i < vocabSize; i++) {
if (logits[i] > maxLogit) {
maxLogit = logits[i];
}
}
// Compute exp and sum
let sum = 0;
for (let i = 0; i < vocabSize; i++) {
probs[i] = Math.exp(logits[i] - maxLogit);
sum += probs[i];
}
// Normalize
for (let i = 0; i < vocabSize; i++) {
probs[i] /= sum;
}
return probs;
}
/**
* Score a candidate move by computing token probabilities
*
* Clones the game, applies the move, generates new TGN, and computes
* the probability of the move tokens.
*/
async scoreMove(game: TrigoGame, move: Move): Promise<number> {
// Clone the game
const clonedGame = game.clone();
// Apply the move to the cloned game
let success: boolean;
if (move.isPass) {
success = clonedGame.pass();
} else if (move.x !== undefined && move.y !== undefined && move.z !== undefined) {
success = clonedGame.drop({ x: move.x, y: move.y, z: move.z });
} else {
// Invalid move format
return -1000;
}
if (!success) {
// Invalid move, return very low probability
return -1000;
}
// Generate TGN from both original and cloned game
const newTGN = clonedGame.toTGN().trim();
// Extract the move substring
// The move should be the new content added after the current TGN
const moveTokens = this.extractMoveTokens(newTGN);
if (moveTokens.length === 0) {
// Could not extract move, return low probability
return -100;
}
// Convert new TGN to tokens
const tokens = this.stringToTokens(newTGN);
// Get configuration
const config = this.inferencer.getConfig();
const seqLen = config.seqLen;
const vocabSize = config.vocabSize;
// Truncate if too long
if (tokens.length > seqLen) {
tokens.splice(0, tokens.length - seqLen);
}
// Run inference (START_TOKEN will be prepended by inferencer)
const logits = await this.inferencer.runInference(tokens);
// Compute probability for the move tokens
// Note: inferencer prepends START_TOKEN, so positions are offset by +1
// Token sequence: [START_TOKEN, ...tokens, PAD, PAD, ...]
// Position in output: token_i is at position i+1 in the padded sequence
// Find where move tokens start in original token sequence
const moveStartInTokens = tokens.length - moveTokens.length;
let logProb = 0;
for (let i = 0; i < moveTokens.length; i++) {
// Position of this move token in the original tokens array
const tokenPos = moveStartInTokens + i;
// Skip if position is out of bounds
// Logits at position tokenPos predict the token at tokenPos+1
if (tokenPos < 0 || tokenPos >= tokens.length) continue;
const offset = tokenPos * vocabSize;
const tokenLogits = logits.slice(offset, offset + vocabSize);
const probs = this.softmax(tokenLogits, vocabSize);
const tokenId = moveTokens[i];
const prob = probs[tokenId];
if (prob > 0) {
logProb += Math.log(prob);
} else {
// If probability is zero, assign very low prob
logProb += -100;
}
}
return logProb;
}
/**
* Extract move tokens from TGN difference
* Returns the tokens that were added between currentTGN and newTGN
*/
private extractMoveTokens(tgn: string): number[] {
const moveCapture = tgn.match(/[Pa-z0]+$/);
return this.stringToTokens(moveCapture ? moveCapture[0] : "");
}
/**
* Select the best move using the language model
*
* Scores all valid moves and returns the one with highest probability (argmax).
*/
async selectBestMove(game: TrigoGame): Promise<Move | null> {
if (!this.isInitialized()) {
throw new Error("Agent not initialized. Pass initialized inferencer to constructor.");
}
console.log("[TrigoAgent] Selecting move...");
// Get current player as string
const currentPlayer = this.stoneToPlayer(game.getCurrentPlayer());
// Get all valid moves
const validMoves: Move[] = game.validMovePositions().map((pos) => ({
x: pos.x,
y: pos.y,
z: pos.z,
player: currentPlayer
}));
validMoves.push({ player: currentPlayer, isPass: true }); // Add pass move
if (validMoves.length === 0) {
console.log("[TrigoAgent] No valid moves available");
return null;
}
console.log(`[TrigoAgent] Evaluating ${validMoves.length} valid moves...`);
// Score each move
const scores: MoveScore[] = [];
for (const move of validMoves) {
const logProb = await this.scoreMove(game, move);
scores.push({
move,
score: Math.exp(logProb), // Convert log prob to probability
logProb
});
}
// Find best move (argmax)
scores.sort((a, b) => b.logProb - a.logProb);
const bestMove = scores[0];
console.debug("scores:", scores);
console.log("[TrigoAgent] Best move:", bestMove.move, "score:", bestMove.score.toFixed(6));
console.log("[TrigoAgent] Top 5 moves:");
for (let i = 0; i < Math.min(5, scores.length); i++) {
console.log(` ${i + 1}. ${scores[i].move}: ${scores[i].score.toFixed(6)}`);
}
return bestMove.move;
}
/**
* Clean up resources
*/
destroy(): void {
this.inferencer.destroy();
console.log("[TrigoAgent] Destroyed");
}
}
|