You signed in with another tab or window. Reload to refresh your session.You signed out in another tab or window. Reload to refresh your session.You switched accounts on another tab or window. Reload to refresh your session.Dismiss alert
Copy file name to clipboardExpand all lines: docs/modding-tools/ai/evaluation-function-constructor.md
+141-4Lines changed: 141 additions & 4 deletions
Display the source diff
Display the rich diff
Original file line number
Diff line number
Diff line change
@@ -59,21 +59,21 @@ Program for creating evaluation functions.
59
59
<tr>
60
60
<td>-c[d]</td>
61
61
<td>
62
-
Convert text to binary data and generate initial patterns. <br />d - Show duplicates being found in text data
62
+
Convert text to binary data and generate initial patterns. <br /><ul>d - Show duplicates being found in text data</ul>
63
63
</td>
64
64
</tr>
65
65
<tr>
66
66
<td>
67
67
-g\{{"c,p\[s],b"}}
68
68
</td>
69
69
<td>
70
-
gc - Generate configurations from atomic features based on test data. <br />gp - generate patterns from configurations being generated. <br />gps - Show all configurations during generation. <br />gb - Generate pattern basis from patterns being generated
70
+
<ul>gc - Generate configurations from atomic features based on test data. <br />gp - generate patterns from configurations being generated. <br />gps - Show all configurations during generation. <br />gb - Generate pattern basis from patterns being generated</ul>
71
71
</td>
72
72
</tr>
73
73
<tr>
74
74
<td>-f[p,u]</td>
75
75
<td>
76
-
Fit weights of pattern configurations. <br />p - Force to use probabilistic weight fitting algorithm. <br />u - Force to use previous parameters if exist
76
+
Fit weights of pattern configurations. <br /><ul>p - Force to use probabilistic weight fitting algorithm. <br />u - Force to use previous parameters if exist</ul>
77
77
</td>
78
78
</tr>
79
79
<tr>
@@ -91,7 +91,7 @@ Program for creating evaluation functions.
91
91
<tr>
92
92
<td>-b[f]</td>
93
93
<td>
94
-
Build evaluation function. <br />f - Save parameters in float (default is double)
94
+
Build evaluation function. <br /><ul>f - Save parameters in float (default is double)</ul>
95
95
</td>
96
96
</tr>
97
97
<tr>
@@ -108,3 +108,140 @@ Program for creating evaluation functions.
108
108
</tr>
109
109
</tbody>
110
110
</table>
111
+
112
+
## Config file
113
+
114
+
Parameters contained in `efc.ini`.
115
+
116
+
### File Names
117
+
118
+
| Key | Description |
119
+
|---|---|
120
+
| LogData | Dual output log file |
121
+
| TextData | Human-readable training data |
122
+
| BinaryData | Binary test data with duplicates removed |
123
+
| ConfigData | Configurations organized by cardinality |
These four parameters control the gradient descent optimization algorithm that learns pattern weights from training data. The algorithm implements momentum-based gradient descent with configurable convergence criteria.
132
+
133
+
| Key | Description |
134
+
|---|---|
135
+
| Epsilon | Convergence Threshold. Defines the minimum per-test-example improvement in squared error required to continue optimization |
136
+
| Alpha | Learning Rate. Scales the gradient magnitude during parameter updates. Larger values take bigger steps in parameter space |
137
+
| Beta | Momentum Coefficient. Weights the previous update direction in the current update, enabling momentum to accelerate convergence and escape shallow local minima. |
138
+
| MaxIterationCount | Safety Limit. Hard upper bound on optimization iterations to prevent infinite loops or excessive runtime |
139
+
140
+
### Probabilistic Weight Fitting Parameters
141
+
142
+
These parameters enable stochastic perturbation during optimization to escape local minima. When active, random subsets of parameters are temporarily frozen while others are updated, creating exploration dynamics.
143
+
144
+
<table><thead>
145
+
<tr>
146
+
<th>Key</th>
147
+
<th>Description</th>
148
+
</tr></thead>
149
+
<tbody>
150
+
<tr>
151
+
<td>RandomFactor</td>
152
+
<td rowspan="2">Exploration Rate. Together determine the fraction of parameters randomly frozen during each update cycle. Each parameter has RandomProbability / RandomFactor chance of being frozen</td>
153
+
</tr>
154
+
<tr>
155
+
<td>RandomProbability</td>
156
+
</tr>
157
+
<tr>
158
+
<td>RandomUpdate</td>
159
+
<td>Refresh Interval. Controls how many iterations use the same random mask before regenerating. Lower values increase variability, higher values provide stability</td>
160
+
</tr>
161
+
<tr>
162
+
<td>RandomStartSeed</td>
163
+
<td>Reproducibility. Initializes the pseudo-random number generator for reproducible experiments. Same seed produces identical random sequences</td>
164
+
</tr>
165
+
</tbody>
166
+
</table>
167
+
168
+
### Configuration Generation Parameters
169
+
170
+
These parameters control the complexity and filtering of patterns generated during the configuration and pattern generation stages. They determine which variable combinations are considered as features for the evaluation function.
171
+
172
+
| Key | Description |
173
+
|---|---|
174
+
| MatchThreshold | Minimum Pattern Frequency. Defines the minimum number of test examples in which a configuration must appear to be considered as a viable pattern. Acts as a frequency filter during configuration generation |
175
+
| MaxCardinality | Pattern Complexity Limit. Limits the maximum number of variables that can be combined in a single pattern. Controls the combinatorial explosion during configuration generation |
176
+
177
+
### Patterns Generation Parameters
178
+
179
+
| Key | Description |
180
+
|---|---|
181
+
| PatternExistanceCoefficient | Coverage Threshold. Filters patterns based on their coverage ratio: count / complexity. A pattern must appear in at least this fraction of its possible instantiations to be kept |
0 commit comments