1 /** Neuroevolution. 2 3 See_Also: https://en.wikipedia.org/wiki/Neuroevolution 4 5 TODO Profile use of cellops and logops in a specific domain or pattern och 6 and use roulette wheel sampling based on these in future patterns. 7 8 TODO reuse bit_traits.packedBitSizeOf 9 10 TODO reuse `IndexedBy` `Cells` and `CellIx` 11 */ 12 module nxt.nevo; 13 14 @safe pure: 15 16 /** Low-Level (Genetic Programming) Operation Network Node Types often implemented 17 in a modern hardware (CPU/GPU). 18 19 See_Also: http://llvm.org/docs/LangRef.html#typesystem 20 See_Also: http://llvm.org/docs/LangRef.html#instref 21 */ 22 enum Lop 23 { 24 summ, /// Sum. 25 prod, /// Product. 26 emin, /// Mininum. 27 emax /// Maximum. 28 } 29 30 /** Obselete. 31 */ 32 enum LOPobseleted 33 { 34 id, /**< Identity. */ 35 36 /* \name arithmetical: additive */ 37 /* @{ */ 38 add, /**< add (mi1o) */ 39 sub, /**< subtract (mi1o) */ 40 neg, /**< negation (1i1o) */ 41 /* @} */ 42 43 /* \name arithmetical: multiplicative */ 44 /* @{ */ 45 mul, /**< multiply (mimo) */ 46 div, /**< divide (mi1o) */ 47 pow, /**< power (2i1o) */ 48 inv, /**< inverse (1i1o) */ 49 /* @} */ 50 51 /* \name Harmonic */ 52 /* @{ */ 53 sin, /**< sine (1i1o) */ 54 cos, /**< cosine (1i1o) */ 55 tan, /**< tangens (1i1o) */ 56 57 /* \name Inverse Harmonic */ 58 csc, /**< 1/sin(a) */ 59 cot, /**< cotangens (1i1o) */ 60 sec, /**< 1/cos(a) */ 61 62 // Arcus 63 acos, /**< arcus cosine (1i1o) */ 64 asin, /**< arcus sine (1i1o) */ 65 atan, /**< arcus tanges (1t2i1o) */ 66 acot, /**< arcus cotangens (1i1o) */ 67 acsc, /**< arcus css (1i1o) */ 68 /* @} */ 69 70 /* \name hyperbolic */ 71 /* @{ */ 72 cosh, sinh, tanh, coth, csch, 73 acosh, asinh, atanh, acoth, acsch, 74 /* @} */ 75 76 // /* \name imaginary numbers */ 77 // /* @{ */ 78 // real, imag, conj, 79 // /* @} */ 80 81 /* \name reorganizer/permutators/permutations */ 82 /* @{ */ 83 identity, /**< identity (copy) (1imo) */ 84 interleave, /**< interleave (1imo) */ 85 flip, /**< flip (1i1o) */ 86 mirror, /**< mirror (1i1o) */ 87 shuffle, /**< shuffle (1i1o) */ 88 /* @} */ 89 90 /* \name generators */ 91 /* @{ */ 92 ramp, /**< linear ramp (2imo) */ 93 zeros, /**< zeros (0imo) */ 94 ones, /**< ones (0imo) */ 95 rand, /**< ranodm (0imo) */ 96 /* @} */ 97 98 /* \name comparison */ 99 /* @{ */ 100 eq, /**< equality (mi1o) */ 101 neq, /**< non-equality (mi1o) */ 102 lt, /**< less-than (2i1o) */ 103 lte, /**< less-than or equal (2i1o) */ 104 gt, /**< greater-than (2i1o) */ 105 gte, /**< greater-than or equal (2i1o) */ 106 /* @} */ 107 108 /* \name logical */ 109 /* @{ */ 110 all, /**< all (non-zero) (mi1o) */ 111 any, /**< any (non-zero) (mi1o) */ 112 none, /**< none (zero) (mi1o) */ 113 /* @} */ 114 } 115 116 version(none) 117 { 118 /** Return non-zero if \p lop is an \em Arithmetical \em Additive Operation. */ 119 bool isAdd(Lop lop) 120 { 121 return (lop == Lop.add || 122 lop == Lop.sub || 123 lop == Lop.neg); 124 } 125 126 /** Return non-zero if \p lop is an \em Arithmetical \em Additive Operation. */ 127 bool isMul(Lop lop) 128 { 129 return (lop == Lop.mul || 130 lop == Lop.div || 131 lop == Lop.pow || 132 lop == Lop.inv); 133 } 134 135 bool isArithmetic(Lop lop) 136 { 137 return isAdd(lop) && isArithmetic(lop); 138 } 139 140 /** Return non-zero if \p lop is an \em Trigonometric Operation. */ 141 bool isTrigonometric(Lop lop) 142 { 143 return (lop == Lop.cos || 144 lop == Lop.sin || 145 lop == Lop.tan || 146 lop == Lop.cot || 147 lop == Lop.csc || 148 lop == Lop.acos || 149 lop == Lop.asin || 150 lop == Lop.atan || 151 lop == Lop.acot || 152 lop == Lop.acsc 153 ); 154 } 155 156 /** Return non-zero if \p lop is an \em Hyperbolic Operation. */ 157 bool isHyperbolic(Lop lop) 158 { 159 return (lop == Lop.cosh || 160 lop == Lop.sinh || 161 lop == Lop.tanh || 162 lop == Lop.coth || 163 lop == Lop.csch || 164 lop == Lop.acosh || 165 lop == Lop.asinh || 166 lop == Lop.atanh || 167 lop == Lop.acoth || 168 lop == Lop.acsch); 169 } 170 171 /** Return non-zero if \p lop is a \em Generator Operation. */ 172 bool isGen(Lop lop) 173 { 174 return (lop == Lop.ramp || 175 lop == Lop.zeros || 176 lop == Lop.ones || 177 lop == Lop.rand); 178 } 179 180 /** Return non-zero if \p lop is a \em Comparison Operation. */ 181 bool isCmp(Lop lop) 182 { 183 return (lop == Lop.eq || 184 lop == Lop.neq || 185 lop == Lop.lt || 186 lop == Lop.lte || 187 lop == Lop.gt || 188 lop == Lop.gte); 189 } 190 191 /** Return non-zero if \p lop is a \em Permutation/Reordering Operation. */ 192 bool isPermutation(Lop lop) 193 { 194 return (lop == Lop.identity || 195 lop == Lop.interleave || 196 lop == Lop.flip || 197 lop == Lop.mirror || 198 lop == Lop.shuffle); 199 } 200 } 201 202 /// Cell Operation. 203 enum CellOp 204 { 205 seqClone, /// sequential clone 206 parClone, /// parallel clone 207 } 208 alias CellOps = Owned!(DynamicArray!CellOp); // TODO bit-pack 209 210 /**m Network (Transformation) Operation Type Code. 211 * 212 * Instructions for a Program that builds \em Computing Networks (for 213 * example Artificial Nerual Networks ANN). 214 * 215 * TODO What does \em nature cell this information bearer: Closest I 216 * have found is http://en.wikipedia.org/wiki/Allele. 217 */ 218 enum Gop 219 { 220 /* \name Structural: Inter-Node */ 221 /* @{ */ 222 seq, /**< Sequential Growth. Inherit Initial Value. Set weight to +1. State is \em ON. */ 223 par, /**< Parallel Growth */ 224 225 clone, /**< Clone. Like \c PAR but */ 226 /* @} */ 227 228 /* \name Local: Intra-Node */ 229 /* @{ */ 230 setLOp, /**< Set Node Logic Operation (see \c Lop). */ 231 232 setOType, /**< Set Out-Type. Either \c int, \c float or \c double for now. */ 233 234 incBias, /**< Increase Threshold */ 235 decBias, /**< Decrease Threshold */ 236 237 /* There is \em no need for "in-precision", since these are the same 238 as in-precision the "previous" nodes in the network. */ 239 incOpRec, /**< Increase Out-Precision */ 240 decOpRec, /**< Decrease Out-Precision */ 241 242 incLinkreg, /**< Increase In-Link-Register */ 243 decLinkreg, /**< Decrease In-Link-Register */ 244 245 on, /**< \em Activate Current In-Link. */ 246 off, /**< \em Deactivate Current In-Link. */ 247 248 /* @} */ 249 250 wait, /**< Wait (Delay) one clock-cycle? */ 251 rec, /**< \em Recurse up to top (decreasing life with one). */ 252 253 jmp, /**< \em Jump to sub-Network/Procedure/Function. */ 254 def, /**< \em Define sub-Network/Procedure/Function. */ 255 256 end, /**< \em Recursion Terminator in serialized code (stream of t's). */ 257 null_, /**< \em Network Terminator in serialized code (stream of t's). */ 258 } 259 260 // wchar 261 // wchar_from_LOP(Lop lop) 262 // { 263 // wchar_t wc = UNICODE_UNKNOWN; 264 // switch (lop) { 265 // case Lop.add: wc = UNICODE_CIRCLED_PLUS; break; 266 // case Lop.sub: wc = UNICODE_CIRCLED_MINUS; break; 267 // case Lop.neg: wc = UNICODE_SQUARED_MINUS; break; 268 // case Lop.mul: wc = UNICODE_CIRCLED_TIMES; break; 269 // case Lop.div: wc = UNICODE_CIRCLED_DIVISION_SLASH; break; 270 // case Lop.pow: wc = UNICODE_UNKNOWN; break; 271 // case Lop.inv: wc = UNICODE_CIRCLED_DIVISION_SLASH; break; 272 // default: break; 273 // } 274 // return wc; 275 // } 276 277 import std.bitmanip : BitArray; 278 import std.random : Random, uniform; 279 280 import nxt.dynamic_array : DynamicArray; 281 import nxt.dynamic_array : DynamicArray; 282 283 import nxt.typecons_ex : IndexedBy; 284 import nxt.owned : Owned; 285 286 import nxt.variant : FastAlgebraic; 287 288 alias Data = FastAlgebraic!(long, double); 289 alias Datas = Owned!(DynamicArray!Data); 290 291 /// Scalar Operation Count. 292 alias OpCount = size_t; 293 294 /// Relative Cell index. 295 alias CellRIx = ptrdiff_t; 296 297 /// Relative Cell indexs. 298 alias CellRIxs = Owned!(DynamicArray!CellRIx); 299 300 /// Calculating Cell. 301 struct Cell 302 { 303 import std.algorithm.iteration : map, filter, fold; 304 305 @safe pure: 306 307 this(Lop lop, size_t inputRIxsLength, size_t cellCount) 308 { 309 } 310 311 /// Randomize a cell. 312 void randomize(Lop lop, size_t inputRIxsLength, size_t cellCount) @trusted 313 { 314 auto gen = Random(); 315 this.lop = lop; 316 317 // TODO use std.algorithm to fill in `inputCellRIxs` 318 inputCellRIxs.length = inputRIxsLength; 319 foreach (ref rix; inputCellRIxs[]) 320 { 321 rix = uniform(cast(CellRIx)0, cast(CellRIx)cellCount/10, gen); 322 } 323 } 324 325 nothrow: 326 327 OpCount execute(const ref Datas ins, ref Datas outs) const 328 { 329 if (ins.empty) { return typeof(return).init; } 330 final switch (lop) 331 { 332 case Lop.summ: return summ(ins, outs); 333 case Lop.prod: return prod(ins, outs); 334 case Lop.emin: return emin(ins, outs); 335 case Lop.emax: return emax(ins, outs); 336 } 337 } 338 339 /// Sum of `ins`. 340 OpCount summ(const ref Datas ins, ref Datas outs) const @trusted 341 { 342 import std.algorithm.iteration : sum; 343 outs.length = 1; 344 outs[0] = ins[].filter!(_ => _.hasValue) 345 .map!(_ => _.commonValue) 346 .sum(); 347 return ins.length - 1; 348 } 349 350 /// Product of `ins`. 351 OpCount prod(const ref Datas ins, ref Datas outs) const @trusted 352 { 353 outs.length = 1; 354 outs[0] = ins[].filter!(_ => _.hasValue) 355 .map!(_ => _.commonValue) 356 .fold!((a, b) => a * b)(cast(Data.CommonType)1.0); 357 return ins.length - 1; 358 } 359 360 /// Minimum of `ins`. 361 OpCount emin(const ref Datas ins, ref Datas outs) const @trusted 362 { 363 import std.algorithm : minElement; 364 outs.length = 1; 365 outs[0] = ins[].filter!(_ => _.hasValue) 366 .map!(_ => _.commonValue) 367 .minElement(+Data.CommonType.max); 368 return ins.length - 1; 369 } 370 371 /// Maximum of `ins`. 372 OpCount emax(const ref Datas ins, ref Datas outs) const @trusted 373 { 374 import std.algorithm : maxElement; 375 outs.length = 1; 376 outs[0] = ins[].filter!(_ => _.hasValue) 377 .map!(_ => _.commonValue) 378 .maxElement(-Data.CommonType.max); 379 return ins.length - 1; 380 } 381 382 Lop lop; /// operation 383 CellRIxs inputCellRIxs; /// relative indexes to (neighbouring) input cells 384 } 385 alias Cells = IndexedBy!(Owned!(DynamicArray!Cell), `Ix`); 386 387 /// Network/Graph of `Cells`. 388 struct Network 389 { 390 @safe pure /*TODO nothrow @nogc*/: 391 392 this(size_t cellCount) 393 { 394 auto gen = Random(); 395 396 cells.reserve(cellCount); 397 temps.reserve(cellCount); 398 399 foreach (immutable i; 0 .. cellCount) 400 { 401 cells ~= Cell(gen.uniform!Lop, 10, cellCount); 402 temps ~= Data(gen.uniform!long); 403 } 404 } 405 406 /// One step forward. 407 OpCount step() @trusted 408 { 409 typeof(return) opCount = 0; 410 411 foreach (immutable i, ref cell; cells) 412 { 413 Datas ins; 414 ins.reserve(cell.inputCellRIxs.length); 415 foreach (immutable rix; cell.inputCellRIxs) 416 { 417 ins ~= temps[(i + rix) % cells.length]; // relative index 418 } 419 420 Datas tempOuts; // data to be filled 421 if (immutable opCount_ = cell.execute(ins, tempOuts)) 422 { 423 opCount += opCount_; 424 temps[i] = tempOuts[0]; 425 } 426 } 427 return opCount; 428 } 429 430 BitArray pack() nothrow @nogc 431 { 432 typeof(return) bits; 433 return bits; 434 } 435 436 static typeof(this) unpack(in BitArray bits) nothrow @nogc 437 { 438 typeof(this) that; 439 return that; 440 } 441 442 Cells cells; /// operation/function cells 443 Datas temps; /// temporary outputs from cells 444 } 445 446 /// Generative Code. 447 struct Code 448 { 449 @safe pure nothrow: 450 451 CellOps cellOps; // cell operations, an indirect generative network encoding 452 alias cellOps this; 453 454 Cells.Ix writerIxs; /// index to writers 455 456 Network generate() const 457 { 458 typeof(return) network; 459 foreach (immutable cellOp; cellOps[]) 460 { 461 } 462 return network; 463 } 464 } 465 466 unittest 467 { 468 immutable cellCount = 1_000; 469 auto task = Network(cellCount); 470 471 immutable stepCount = 1_0; 472 foreach (immutable i; 0 .. stepCount) 473 { 474 task.step(); 475 } 476 477 // dbg("DONE"); 478 }