I thought I would open a discussion on hacking the files in data/neuralnet/
I'll start off with nnettackle.dat because I want to get the game to be a bit rougher.
1st, Neural nets are basically a way to extrapolate complex decisions given a limited amount of information. A data file for a neural net should contain numeric data points which the program will use to calculate output behaviour based on input information. There are three parts of a neural net. INPUT LAYER, CALCULATION MATRIX, and OUTPUT LAYER. If we interpret the nnettackle.dat as a little endian file containing 32-bit float numbers, we observe some patterns:
Section 1: ¿¿ Input layer parameters ??
The first five numbers are exact integers: 4.0, 7.0, 10.0, 10.0, 1.0
Section 2: ¿¿ Computational layer matrix ??
The next 201 numbers are non-integers valued between -7.5 and +7.5
Section 3: Low Range scale of calculation matrix output updated
Section3[0]= -11.204
Section3[1]= -102.513
Section3[0] Scales the probability of a tackle using an aggressive or poor tackle animation. At the default value of -11.204, these animations are rare. When increased to a range from -100.00000 to -200.00000, these animation sequences are more frequent. I have not determined the purpose of Section3[1] yet.
Section 4: High Range scale of calculation matrix output updated
Section4[0]= 94.148003
Section4[1]= 118.445
Section3[0] and Section4[0] work together to define the range of output from the calculation matrix, so default range is -11.204 <--> 94.148003
Decreasing Section4[0] will decrease the probability that a clean tackle animation sequence will be used. I have not determined the purpose of Section4[1] yet.
Section 5: ¿¿ Output layer parameters ??
The last five numbers are exact integers: 1.0, 180.0, 180.0, 1.0, 98.0
Changing these values has no effect in my test so far.
CHRISTMAS UPDATE
Link: Improved Tackle behavior version 0.1
Changelog(v0.1): Decreased Section3[0] 2-fold to -205.026001
Changelog(v0.1): Decreased Section4[0] by 1/2 to 49.074001
I'll start off with nnettackle.dat because I want to get the game to be a bit rougher.
1st, Neural nets are basically a way to extrapolate complex decisions given a limited amount of information. A data file for a neural net should contain numeric data points which the program will use to calculate output behaviour based on input information. There are three parts of a neural net. INPUT LAYER, CALCULATION MATRIX, and OUTPUT LAYER. If we interpret the nnettackle.dat as a little endian file containing 32-bit float numbers, we observe some patterns:
Section 1: ¿¿ Input layer parameters ??
The first five numbers are exact integers: 4.0, 7.0, 10.0, 10.0, 1.0
Section 2: ¿¿ Computational layer matrix ??
The next 201 numbers are non-integers valued between -7.5 and +7.5
Section 3: Low Range scale of calculation matrix output updated
Section3[0]= -11.204
Section3[1]= -102.513
Section3[0] Scales the probability of a tackle using an aggressive or poor tackle animation. At the default value of -11.204, these animations are rare. When increased to a range from -100.00000 to -200.00000, these animation sequences are more frequent. I have not determined the purpose of Section3[1] yet.
Section 4: High Range scale of calculation matrix output updated
Section4[0]= 94.148003
Section4[1]= 118.445
Section3[0] and Section4[0] work together to define the range of output from the calculation matrix, so default range is -11.204 <--> 94.148003
Decreasing Section4[0] will decrease the probability that a clean tackle animation sequence will be used. I have not determined the purpose of Section4[1] yet.
Section 5: ¿¿ Output layer parameters ??
The last five numbers are exact integers: 1.0, 180.0, 180.0, 1.0, 98.0
Changing these values has no effect in my test so far.
CHRISTMAS UPDATE
Link: Improved Tackle behavior version 0.1
Changelog(v0.1): Decreased Section3[0] 2-fold to -205.026001
Changelog(v0.1): Decreased Section4[0] by 1/2 to 49.074001