Working With DRC’s Parameters

Introduction

DRC’s parameters are used to control many aspects of the behaviour of the model during simulations. The purpose of each of the parameters is described below.

Parameter names are formed by concatenating a term for each model layer that the parameter is associated with and a term for the function it is associated with. LetterLateralInhibition, for example, is associated with the letter layer and the lateral inhibition function.

The terms used in parameter names to refer to each model layer are listed in the table below:

Layer                   Term
-------------------------------
Visual feature layer    Feature
Letter layer            Letter
Orthographic lexicon    Orthlex
Phonological lexicon    Phonlex
GPC Route               GPC
Phoneme layer           Phoneme

Parameter Value Ranges

Most parameters should be set to a value between 0 and 1. When this is not the case it will be clearly stated in the parameter description. All parameter values should be positive.

Default Values

The default value of each parameter is defined on a per-language basis. You can view the default parameter values for a language by opening the “default.parameters” file, found inside the language folder, in a text editor. Suitable editors include Notepad in Windows, or the TextEdit application in Mac OS X.

Any parameter not mentioned in the default.parameters file has a default value of zero. You should not modify the language default.parameters file unless your intention is to create a new language for DRC.

Finding Out Which Parameter Values Were Used In a Simulation

DRC stores the values that the parameters had at the beginning of a job in a file named “parameters” in the output folder. Any changes to parameter values that occurred during the job are reported in the RT file.

Changing Parameter Values

There are three methods for changing a parameter from its default value for a simulation. You may use one or more of these methods simultaneously.

You can specify them on the command line with the ‘-P’ (uppercase P) option. Parameter values set in this way continue to apply throughout the job, unless the parameter is changed again via a batch command or parameter stepping. This example command line shows how to set the FeatureLetterExcitation parameter to 0.1:
```
drc -P FeatureLetterExcitation 0.1 ...
```
You can change multiple parameters in this way. For example:
```
drc -P FeatureLetterExcitation 0.1 -P GPCOnset 10 ...
```
If the same parameter is changed more than once in this fashion, the value specified latest on the command line will be used.
You can create a new file in a text editor, listing parameter names and their values. Each parameter should appear on a separate line, followed by a space, and then the value you want the parameter to have. For example:
```
FeatureLetterExcitation 0.1
GPCOnset 10
```
You can then tell DRC to use the values in the file with the ‘-p’ (lowercase p) option. For example:
```
drc -p parameters.txt ...
```
You don’t need to provide a value for every single parameter in your parameters file. Any parameter not listed may be changed using one of the other methods, or it will have its default value.

If you attempt to use a parameters file and parameter values specified on the command line simultaneously, DRC will load the parameters from the file first and then apply any changes specified on the command line.
You can change parameters in a batch file by adding lines like the following:
```
! FeatureLetterExcitation 0.1
! GPCOnset 10
```
The parameter value is not changed until the line containing the change is reached. This means that you can create a batch file that uses different parameter values for different sets of stimuli. In the following example batch file, the word TREE will be run 3 times. On the first run the FeatureLetterExcitation parameter will have its default value. On the second it will have a value of 0.2, and on the third it will have a value of 0.3:
```
TREE
! FeatureLetterExcitation 0.2
TREE
! FeatureLetterExcitation 0.3
TREE
```

Parameter Stepping

DRC allows you to quickly see the results of stepping through a number of sequential values for a parameter via the parameter stepping function.

There are two ways to use the parameter stepping function, selected by using either the ‘-S’ or ‘-S2′ command line switches. In the first, you specify the starting value for the parameter, the ending value and the number of steps that should be taken. The simulation will be repeated as many times as was specified. On the first run the parameter will have the starting value and on subsequent runs the parameter will increase by equal amounts so that on the last run the parameter will have the ending value.

In the following example the simulation is repeated 3 times, with FeatureLetterExcitation starting at 0.1, and ending at 0.3. On the second run the value of FeatureLetterExcitation will be 0.2:

drc -S FeatureLetterExcitation 0.1 0.3 3 ...

In the second method, you specify the starting value, ending value and the amount by which the parameter should be increased on each step. The simulation will be repeated as many times as possible without allowing the parameter to exceed the ending value.

In the following example the simulation will be run twice, with FeatureLetterExcitation having a value of 0.1 on the first run and 0.25 on the second:

drc -S2 FeatureLetterExcitation 0.1 0.3 0.15 ...

Parameter stepping may be used with batch files. The entire batch file will be run for each parameter value.

Changing Parameters Mid-Simulation

By using the “AUTORESET OFF” batch command, it is possible to run 1 or more cycles of a simulation with one set of parameter values, and then run more cycles with a new set of values.

In the following example batch file the input TREE is run for 10 cycles before the ActivationRate parameter is changed to 0.3 and a further 10 cycles are run:

! AUTORESET OFF
10 TREE
! ActivationRate 0.3
10 TREE

General Parameters

ActivationRate

“The activation rate parameter is used to scale the speed at which the activation of the unit changes with respect to the difference between its current activation value and the scaled net input. Small values cause the activation level to change more slowly; high values cause the activation level to change more quickly (Coltheart et al., 2001).”

FrequencyScale

This parameter controls the dominance of the frequency effect in the orthographic and phonological lexicons.

MinReadingPhonology

DRC decides whether a stimulus has been named by examining the phoneme sets in left-to-right order. If each phoneme set, up to and including the first set with the blank phoneme as the most active phoneme, has a phoneme with an activation level greater than or equal to the MinReadingPhonology then the stimulus has been named.

Excitatory and Inhibitory Link Strength Parameters

FeatureLetterExcitation
FeatureLetterInhibition

LetterOrthlexExcitation
LetterOrthlexInhibition

OrthlexPhonlexExcitation
OrthlexPhonlexInhibition
OrthlexLetterExcitation
OrthlexLetterInhibition

PhonlexPhonemeExcitation
PhonlexPhonemeInhibition
PhonlexOrthlexExcitation
PhonlexOrthlexInhibition

PhonemePhonlexExcitation
PhonemePhonlexInhibition

GPCPhonemeExcitation

Each excitatory and inhibitory link between layers in the DRC model has a strength parameter that controls the amount of excitation or inhibition that is contributed across that link.

The parameters that control these strengths are named by specifying the term for the layer that the excitation or inhibition is coming from followed by the term for the layer that the excitation or inhibition is going to followed by either Excitation or Inhibition.

The FeatureLetterExcitation parameter, for example, controls the strength of the excitatory link from the visual feature layer to the letter layer. The PhonlexOrthlexInhibition parameter, on the other hand, controls the strength of the inhibitory link from the phonological lexicon to the orthographic lexicon.

Lateral Inhibition Parameters

LetterLateralInhibition
OrthlexLateralInhibition
PhonlexLateralInhibition
PhonemeLateralInhibition

The strength of lateral inhibition within layers is controlled by these parameters.

GPC Route Parameters

GPCOnset controls the number of cycles that pass before the GPC route begins to operate. This parameter may be set to an integer value greater than, or equal to, zero.

GPCCriticalPhonology controls the conditions under which new letters become available to the GPC route. Initially only the first letter is available. On each cycle the right-most phoneme set that received excitation from the GPC route on the previous cycle is checked to see if it contains a phoneme with an activation level greater than or equal to GPCCriticalPhonology. If it does, the next letter is made available to the GPC route.

Noise Parameters

LetterNoise
OrthlexNoise
PhonlexNoise
GPCNoise*
PhonemeNoise

The amount of noise added to the net-inputs in each layer is determined by an algorithm that produces a stream of pseudo-random numbers with a normal distribution about a mean of zero. These parameters control the standard-deviation of the stream.

A value of zero will mean that no noise is added to the net-inputs for that layer.

* The GPCNoise parameter controls the level of noise added to the amount of activation the GPC route contributes to the phonemes.

Decay Parameters

LetterDecay
OrthlexDecay
PhonlexDecay
PhonemeDecay

Each cycle the activation levels of the units in each layer are reduced by a proportion of their activation level at the end of the previous cycle. The decay parameters control how large the proportion is. A value of 0 therefore indicates that no reduction occurs.

There is also a function called “unsupported decay”. Unsupported decay works in the same way as normal decay but only applies to units that have a net-input of zero or less.

Unsupported decay is normally only used for the phoneme layer, but can be used with the letter, phoneme and lexical layers as well.

The unsupported decay parameters are:

LetterUnsupportedDecay
OrthlexUnsupportedDecay
PhonlexUnsupportedDecay
PhonemeUnsupportedDecay

Advanced Topics

The following parameters should only be employed by advanced users. They control experimental features of the software.

FeatureDuration

For a set number of cycles, specified by the FeatureDuration parameter, from the start of a simulation, and from any point where the model’s input is changed midway through a simulation, there will be no excitation or inhibition from the visual feature layer to the letter layer.

Threshold Parameters

LetterThreshold
OrthlexThreshold
PhonlexThreshold
PhonemeThreshold

These parameters control the threshold for each layer. Any unit with an activation level lower than the threshold for its layer does not contribute excitation or inhibition to any other unit. This includes lateral inhibition.

GPC Threshold Parameters

GPCInputThreshold applies to the letters in the input for the GPC route. Any letter with an activation level below the threshold is treated as having an activation level of 0 for the purpose of calculating the amount of activation the GPC route contributes to phonemes produced by rules that use that letter.

GPCOutputThreshold applies to the phonemes that the GPC route contributes excitation to. If the amount of excitation that the GPC route contributes to a phoneme is less than this threshold, the GPC route contributes no excitation to that phoneme.

GPCOnsetThreshold is used in the procedure for determining when the onset of the GPC route occurs. If this threshold is greater than zero then the onset of the GPC route is delayed until every set of units in the letter layer contains a letter with an activation level greater than or equal to the threshold. The GPCOnset cycle must still be reached before onset will occur, however.

The GPC Route Crapshoot

GPCEarlyOnsetProbability can be used to specify the probability that onset of the GPC route will occur early. On every cycle before the onset of the GPC route occurs, a random number between 0 and 1 is generated. If this number is greater than or equal to the value of this parameter onset of the GPC route will occur immediately. A value of 0 for this parameter means that there is no possibility of early onset occurring.

GPCEarlyLetterProbability can be used to specify the probability that a letter will become available to the GPC route earlier than it normally would. On every cycle in which a new letter is not made available to the GPC route, a random number is generated. If this number is greater than or equal to the value of this parameter the letter will be made available immediately. A value of 0 for this parameter means that there is no possibility of early letter availability occurring.

GPC Body Rules

GPCRhymePhonemeExcitation controls the strength of the excitation the GPC route contributes to phonemes produced by rules of the “body” type.

References

Coltheart, M., Rastle, K., Perry, C., Langdon, R. & Ziegler, J. (2001). DRC: A Dual Route Cascaded model of visual word recognition and reading aloud. Psychological Review, 108, 204 – 256.

Contents