/******************************************************************************/ /* */ /* CONST.H - System and program limitation constants */ /* This also contains typedefs, structs, et cetera. */ /* */ /* See the comment above BAD_COMPILER. */ /* */ /* The #if above MALLOC controls whether or not the diagnostic memory */ /* allocation routines are used. They only slow things a tiny bit. */ /* */ /* RANDMAX may be system dependent. See your documentation. */ /* */ /* MAX_INPUTS, MAX_HIDDEN and MAX_OUTPUTS are primarily to simplify the */ /* code involved in memory allocation. The defaults provided should be */ /* adequate for any project done on a 16 bit machine. Users of larger */ /* machines will want to increase these a lot. */ /* This is controlled with VERSION_16_BIT. */ /* */ /* TRAIN_BUF_SIZE makes training set memory allocation faster by avoiding */ /* many calls to realloc. Users of 32 bit machines with much memory may */ /* want to increase this considerably. */ /* */ /* Copyright (c) 1993 by Academic Press, Inc. */ /* */ /* All rights reserved. Permission is hereby granted, until further notice, */ /* to make copies of this diskette, which are not for resale, provided these */ /* copies are made from this master diskette only, and provided that the */ /* following copyright notice appears on the diskette label: */ /* (c) 1993 by Academic Press, Inc. */ /* */ /* Except as previously stated, no part of the computer program embodied in */ /* this diskette may be reproduced or transmitted in any form or by any means,*/ /* electronic or mechanical, including input into storage in any information */ /* system for resale, without permission in writing from the publisher. */ /* */ /* Produced in the United States of America. */ /* */ /* ISBN 0-12-479041-0 */ /* */ /******************************************************************************/ /* The author is aware of at least one big-name C++ compiler which is unable to correctly handle pure virtual functions. The following toggle causes the Network destructor to be impure. It is always correct to do this, but bad style. Purists will want to change this to 0, or maybe even remove its code from CLASSES.H and NETWORK.CPP. */ #define BAD_COMPILER 1 /* These interpose memory allocation checking. */ #if 0 #define MALLOC memalloc #define FREE memfree #define REALLOC memrealloc #define MEMTEXT memtext #define MEMCLOSE memclose #else #define MALLOC malloc #define FREE free #define REALLOC realloc #define MEMTEXT notext #define MEMCLOSE nomemclose #endif /* #define VERSION_16_BIT 1 */ #define VERSION_16_BIT 0 #if VERSION_16_BIT #define TRAIN_BUF_SIZE 16384 /* Alloc this much tsets mem (max) at a time */ #define MAX_INPUTS 64 /* Number of network inputs */ #define MAX_HIDDEN 64 /* Number of hidden neurons */ #define MAX_OUTPUTS 64 /* Number of network outputs */ #else #define TRAIN_BUF_SIZE 65536 #define MAX_INPUTS 1024 #define MAX_HIDDEN 1024 #define MAX_OUTPUTS 1024 #endif #define RANDMAX 32767 /* rand() returns from 0 through this */ #define CONTROL_LINE_LENGTH 255 /* Max length of user's control input */ #define MAX_CONTROL_FILES 16 /* Nesting of control files */ #define LTAB_LENGTH 100 /* Activation function table length */ #define LTAB_MAX 10.0 /* Maximum in that table */ #define NEURON_ON 0.9 /* Target output activation levels */ #define NEURON_OFF 0.1 /* must not be forced to 0 and 1 */ /* These are network model codes. If additional networks are defined, they should be appended, leaving existing ones with the same codes, in order to avoid disk file incompatibilites. They must be positive (-1 = unset). */ #define NETMOD_LAYER 1 #define NETMOD_KOH 2 /* These are output model codes. If additional outputs are defined, they should be appended, leaving existing ones with the same codes, in order to avoid disk file incompatibilites. They must be positive (-1 = unset). */ #define OUTMOD_CLASSIFY 1 #define OUTMOD_AUTO 2 #define OUTMOD_GENERAL 3 /* The annealing parameters have a zero suffix for the value used for finding starting weights. The non-zero parameters are for when annealing is used to (hopefully) escape from local minima during learning. */ struct AnnealParams { int temps0 ; // Number of temperatures int temps ; int iters0 ; // Iterations per temperature int iters ; int setback0 ; // Set back iteration counter if improvement int setback ; double start0 ; // Starting temperature double start ; double stop0 ; // Stopping temperature double stop ; } ; struct GenInitParams { int pool ; // Number of individuals in population pool int gens ; // Number of generations after initial population int climb ; // Use hill climbing (elitism)? double overinit ; // Overinitialization factor ( double pcross ; // Probability of crossover double pmutate ; // Probability of mutation } ; struct LearnParams { int init ; // 0=no init, 1=ann noreg, 2=anneal, 3=genetic, 4=regress double quit_err ; // Quit if mean square error fraction of max this low int retries ; // Quit after this many random retries struct AnnealParams *ap ; struct GenInitParams *gp ; struct KohParams *kp ; } ; struct KohParams { // Both network and learning params here int normalization ; // 0=multiplicative, 1=Z int learn_method ; // 0=additive, 1=subtractive double rate ; // Learning rate (0-1) double reduction ; // Learning rate reduction per pass (0-1) } ;
