About the generated Python code

[Polarisamoon]

After converting the network into formulas or Python code, I noticed that some hidden neurons are being referenced before they are computed, which is quite strange. For example:

h[0] = 0.171989*h[31] + 3.169343*i[0] - 0.266712

Here, h[31] has not been calculated yet, so it will use its initial value of 0. Is this normal?

Python Code:
h = np.zeros(35)
o = np.zeros(2)
h[0] = 0.171989*h[31] + 3.169343*i[0] - 0.266712
h[1] = -sin(0.112585*h[28] + 2.851631*i[0] - 2.619233*i[1] - 0.026671)
h[2] = -tanh(1.154439*i[0] + 1.897031*i[1] + 0.352859)
h[3] = nan
h[4] = 1.742897*h[0] - 0.302724*h[9] + 0.276833
h[5] = -0.297971*h[15] - 0.757493*h[24] + 2.0499*i[1] + 0.916681
h[6] = -tanh(2.076261*h[23]*i[1] - 1.89432)
h[7] = -2.081779*i[0] - 1.563756*i[1] + 0.980241
h[8] = 0.156715*h[17] - 1.694074*i[0] + 1.365514*i[1] + 1.311888
h[9] = tanh(0.619051*h[18] + 0.330781)
h[10] = 0.816805*h[11]
h[11] = -sin(0.300976*h[2] + 0.627619)
h[12] = 0.34393 - 0.560032*h[33]
h[13] = tanh(0.00587*h[17]*h[26]*i[1] + 0.874081)
h[14] = Max(0, 0.415834*i[0] + 0.395707*i[1] + 1.267994)
h[15] = Piecewise((-0.703819*i[1] - 0.186702, 0.703819*i[1] + 0.186702 < 0), (-0.003519*i[1] - 0.000934, True))
h[16] = -1.23256*i[0] + 0.079172*i[1] - 1.614457
h[17] = Max(0, 0.065526*h[18] + 0.09237*h[19] + 0.128332*i[1] + 0.361111)
h[18] = 0.432082*h[1] + 0.151905*h[26] + 0.444149*i[0] - 0.605303*i[1] + 0.5776
h[19] = Piecewise((0.913525*i[0] + 0.164543, 0.913525*i[0] + 0.164543 > 0), (0.004568*i[0] + 0.000823, True))
h[20] = -sin(0.129024*h[15] + 0.163934)
h[21] = 1.007134
h[22] = -tanh(-0.495377*h[15] + 0.06355*h[25] + 0.633176)
h[23] = Max(0, 0.715448*h[19] - 0.11983*h[27] - 0.359031)
h[24] = tanh(1.849501*i[0] + 0.141925)
h[25] = Max(0, 0.411862*h[26]*i[0] + 0.304906)
h[26] = 0.172775
h[27] = Max(0, 1.272269*h[24] + 0.399686)
h[28] = -0.110427
h[29] = -0.000292
h[30] = Piecewise((0.59178*h[10] - 0.154874*h[25] - 0.082747, -0.59178*h[10] + 0.154874*h[25] + 0.082747 < 0), (0.002959*h[10] - 0.000774*h[25] - 0.000414, True))
h[31] = Piecewise((0.856304*i[1] + 0.125001, 0.856304*i[1] + 0.125001 > 0), (0.004282*i[1] + 0.000625, True))
h[32] = Max(0, 0.781773*h[28] - 0.127448)
h[33] = tanh(1.0*h[2] + 0.008956)
h[34] = 0.891857*h[8]
o[0] = tanh(0.121949*h[1] + 0.020327*h[12] + 0.031256*h[13] + 0.070328*h[14] + 0.006741*h[17] + 0.030249*h[20] + 0.069293*h[4] - 0.083061*h[5] + 0.146785*h[6] - 0.09403*h[7] - 0.281119*h[8] + 0.310527)
o[1] = tanh(0.071067*h[0] + 0.093137*h[1] + 0.167901*h[13] + 0.038465*h[14] + 0.00907*h[17] + 0.064836*h[22] + 0.116218*h[32] - 0.0562*h[34] + 0.269817*h[5] - 0.012843*h[6] - 0.146425*h[7] - 0.115288)

[WLS2002]

That is so strange. Please provide the running config for me to test.

[Polarisamoon]

Here is my config file:

import os
os.environ["CUDA_VISIBLE_DEVICES"] = "4"
import jax.numpy as jnp

from tensorneat.pipeline import Pipeline
from tensorneat.algorithm.neat import NEAT
from tensorneat.genome import DefaultGenome, BiasNode, DefaultMutation
from tensorneat.problem.func_fit import CustomFuncFit
from tensorneat.common import ACT, AGG
from tensorneat.problem.func_fit import custom2
import jax
import pickle

if __name__ == "__main__":

    pipeline = Pipeline(
        algorithm=NEAT(
            pop_size=1000,
            species_size=20,
            survival_threshold=0.03,
            min_species_size=2,
            genome=DefaultGenome(
                num_inputs=2,
                num_outputs=2,
                max_nodes=100,
                max_conns=250,
                init_hidden_layers=(50,),
                mutation=DefaultMutation(
                    node_add=0.8,  # default is 0.1
                    node_delete=0.4,
                    conn_add=0.8,  # default is 0.2
                    conn_delete=0.4,
                ),
                node_gene=BiasNode(
                    activation_options=[ACT.lelu, ACT.tanh, ACT.sin, ACT.relu,ACT.identity],
                    aggregation_options=[AGG.sum, AGG.product, AGG.mean],
                ),
                output_transform=ACT.tanh,
            ),
        ),
        problem=custom2(),
        generation_limit=100000,
        fitness_target=-1e-6,
        seed=45,
        is_save=True
    )

    # initialize state
    state = pipeline.setup()
    # run until terminate
    state, best = pipeline.auto_run(state)
    # show result
    pipeline.show(state, best)

    # visualize the best individual
    network = pipeline.algorithm.genome.network_dict(state, *best)
    print(pipeline.algorithm.genome.repr(state, *best))
    pipeline.algorithm.genome.visualize(network, save_path="./xor_network_41.svg")

    # best is a tuple (nodes, conns) of jax array
    best = jax.device_get(best)  # convert best to numpy array on cpu
    with open("best.pkl", "wb") as f:
        pickle.dump(best, f)

    # Convert to mathematical formula
    from tensorneat.common.sympy_tools import to_latex_code, to_python_code

    sympy_res = pipeline.algorithm.genome.sympy_func(
        state,
        network,
        sympy_output_transform=ACT.obtain_sympy(ACT.tanh)
    )

    print("\nLatex Formula:")
    latex_code = to_latex_code(*sympy_res)
    print(latex_code)

    print("\nPython Code:")
    python_code = to_python_code(*sympy_res)
    print(python_code)

Here is the best individual saved after 514 iterations. I also found that it exhibits the same phenomenon:
h[3] = -sin(0.580582*h[22]*i[0]*i[1] - 0.469189)：

Python Code:
h = np.zeros(75)
o = np.zeros(2)
h[0] = Max(0, -0.393324*i[1] - 0.304097)
h[1] = -tanh(0.520576*i[1] - 0.621616)
h[2] = Max(0, -0.595404*i[0]*i[1] - 0.476)
h[3] = -sin(0.580582*h[22]*i[0]*i[1] - 0.469189)
h[4] = tanh(0.303605*i[1] + 1.490881)
h[5] = 0.540493*h[16] - 0.527148
h[6] = -sin(1.456117*h[28] + 0.907768*h[42] + 1.714789)
h[7] = -sin(0.897136*h[50] - 0.793472*h[60] + 0.56405*i[0] - 0.260179)
h[8] = Piecewise((-0.207455*i[0]*i[1] + 1.305145, 0.207455*i[0]*i[1] - 1.305145 < 0), (-0.001037*i[0]*i[1] + 0.006526, True))
h[9] = Piecewise((2.485227*h[21] + 0.562212*i[0] + 0.153083, 2.485227*h[21] + 0.562212*i[0] + 0.153083 > 0), (0.012426*h[21] + 0.002811*i[0] + 0.000765, True))
h[10] = Max(0, 0.100386*h[45] - 0.374037)
h[11] = -1.455953*i[0] + 0.32391*i[1] + 0.03569
h[12] = -sin(1.031857*i[0]*i[1] + 0.434128)
h[13] = -0.029788*h[18] + 4.036304*i[0] + 1.445527*i[1] - 0.86662
h[14] = 2.668357
h[15] = sin(0.058248*h[63] + 0.025396)
h[16] = Piecewise((1.602*i[1] - 0.355662, 1.602*i[1] - 0.355662 > 0), (0.00801*i[1] - 0.001778, True))
h[17] = Piecewise((0.246702*h[45] - 0.087146*i[0] + 2.860729*i[1] + 1.412821, 0.246702*h[45] - 0.087146*i[0] + 2.860729*i[1] + 1.412821 > 0), (0.001234*h[45] - 0.000436*i[0] + 0.014304*i[1] + 0.007064, True))
h[18] = tanh(0.047125*h[34] + 0.534341*h[62] + 1.906155)
h[19] = Piecewise((-0.43377*h[59] + 0.478223*i[0] - 0.824897*i[1] - 0.637511, 0.43377*h[59] - 0.478223*i[0] + 0.824897*i[1] + 0.637511 < 0), (-0.002169*h[59] + 0.002391*i[0] - 0.004124*i[1] - 0.003188, True))
h[20] = Piecewise((0.10993*i[0]*i[1] + 3.083846, 0.10993*i[0]*i[1] + 3.083846 > 0), (0.00055*i[0]*i[1] + 0.015419, True))
h[21] = Max(0, 1.13565*i[1] + 0.569425)
h[22] = Piecewise((1.061129*i[1] + 0.941416, 1.061129*i[1] + 0.941416 > 0), (0.005306*i[1] + 0.004707, True))
h[23] = sin(0.516021*h[46] + 0.112596)
h[24] = Max(0, 0.128278*h[37] - 0.668598*h[61] - 0.326777*i[0] + 0.351653)
h[25] = Piecewise((-0.15701*h[26] + 0.286644*h[55] + 1.529387*i[0] - 0.390395, 0.15701*h[26] - 0.286644*h[55] - 1.529387*i[0] + 0.390395 < 0), (-0.000785*h[26] + 0.001433*h[55] + 0.007647*i[0] - 0.001952, True))
h[26] = sin(0.282044*h[29] - 0.59717*h[39] - 0.009626*i[1] + 0.116546)
h[27] = -2.448874*i[0] + 1.027433*i[1] + 0.020469
h[28] = sin(1.571725*i[1] - 0.073324)
h[29] = Piecewise((1.13804*i[0] - 0.587174*i[1] - 0.079896, -1.13804*i[0] + 0.587174*i[1] + 0.079896 < 0), (0.00569*i[0] - 0.002936*i[1] - 0.000399, True))
h[30] = -sin(0.642609*h[1] + 0.182434*i[0] - 1.507243)
h[31] = -0.355151*i[0] - 1.823967*i[1] - 0.174811
h[32] = 1.426471*h[74] - 1.467207*i[1] + 1.084125
h[33] = Max(0, 0.242858*h[53] + 0.178815*i[0] - 0.347632)
h[34] = -tanh(0.030547*h[53] + 0.220346*i[0] + 0.453322*i[1] - 0.189637)
h[35] = Max(0, -0.11602*i[0] + 0.435637*i[1] + 1.69347)
h[36] = tanh(0.480057*i[1] + 0.673986)
h[37] = -sin(0.008656*h[15]*i[0] + 0.31275)
h[38] = sin(0.547461*h[18] + 0.756188)
h[39] = -tanh(0.445631*i[0] + 1.345173)
h[40] = Piecewise((-0.063427*i[0] - 1.473829*i[1] + 1.399987, 0.063427*i[0] + 1.473829*i[1] - 1.399987 < 0), (-0.000317*i[0] - 0.007369*i[1] + 0.007, True))
h[41] = 1.138754*h[44]
h[42] = -0.566675*h[69] + 1.201686*i[0] - 0.659911
h[43] = 0.173262*i[0] + 1.062508*i[1] + 0.003131
h[44] = -3.692652*i[0] + 2.149998*i[1] - 0.960821
h[45] = Max(0, 0.04869*h[14]*h[35]*i[0]*i[1] + 1.542186)
h[46] = Max(0, 0.955812*h[38] + 0.311495*h[43] - 0.320291*h[57] + 0.095745)
h[47] = Max(0, 1.184057*i[0]*i[1] + 1.205099)
h[48] = 0.030837*h[0] - 0.034122*h[11] - 0.0403*h[12] + 0.084885*h[13] + 0.012358*h[2] - 0.03802*h[20] + 0.004205*h[24] + 0.010416*h[25] - 0.04639*h[26] - 0.039131*h[27] + 0.062182*h[29] + 0.008632*h[3] + 0.044204*h[30] + 0.022793*h[31] + 0.033977*h[32] + 0.010851*h[33] + 0.040114*h[35] + 0.029336*h[36] - 0.011805*h[4] + 0.018999*h[40] + 0.012516*h[42] - 0.036528*h[43] - 0.099854*h[44] + 0.009396*h[45] - 0.018849*h[47] - 0.040753*h[6] - 0.008106*h[7] - 0.002093*h[8] + 0.011688*h[9] - 0.006355*o[0] - 0.097498
h[49] = 0.034605*h[0] + 0.002957*h[10] - 0.017802*h[11] + 0.009935*h[12] + 0.034486*h[13] + 0.064219*h[17] - 0.051131*h[19] + 0.015014*h[2] + 0.015692*h[25] - 0.023338*h[26] - 0.076808*h[29] - 0.052635*h[3] + 0.034495*h[30] - 0.042703*h[31] + 0.024121*h[33] - 0.036367*h[34] - 0.046071*h[35] + 0.020079*h[36] + 0.028127*h[37] + 0.05134*h[4] - 0.062171*h[40] - 0.010527*h[42] + 0.061489*h[43] + 0.000863*h[47] + 0.01156*h[51] + 0.041673*h[54] - 0.0104*h[58] + 0.00494*h[6] - 0.043513*h[64] + 0.016655*h[68] - 0.046844*h[7] - 0.012324*h[70] + 0.018825*h[72] - 0.032468*h[8] + 0.122626*h[9] + 0.027858*o[0] - 0.01174*o[1] - 0.05737
h[50] = 1.221238*i[1] - 0.079136
h[51] = sin(1.595441*h[27] - 0.666129)
h[52] = sin(0.232478*i[0] + 0.302679)
h[53] = tanh(0.847736*i[1] + 0.705066)
h[54] = -sin(1.556384*h[73] - 1.454493)
h[55] = Max(0, 0.051253 - 0.48735*i[1])
h[56] = Piecewise((0.56021*h[14] - 0.50237*h[22] + 0.130264*i[0] + 1.566567, 0.56021*h[14] - 0.50237*h[22] + 0.130264*i[0] + 1.566567 > 0), (0.002801*h[14] - 0.002512*h[22] + 0.000651*i[0] + 0.007833, True))
h[57] = 0.639292
h[58] = Max(0, -0.178051*h[32] - 0.782553)
h[59] = Piecewise((-0.076266*h[39]*i[1] - 1.044595, 0.076266*h[39]*i[1] + 1.044595 < 0), (-0.000381*h[39]*i[1] - 0.005223, True))
h[60] = Piecewise((0.110081*h[59] - 0.098415, 0.110081*h[59] - 0.098415 > 0), (0.00055*h[59] - 0.000492, True))
h[61] = Piecewise((-0.770219*h[60] - 0.565254, 0.770219*h[60] + 0.565254 < 0), (-0.003851*h[60] - 0.002826, True))
h[62] = 0.619035*h[56] + 0.128363*h[67] + 0.217896
h[63] = Max(0, 0.705778*i[1] - 0.004013)
h[64] = 0.774546*h[71] - 0.385223
h[65] = sin(1.895392*i[1] + 0.172721)
h[66] = Piecewise((0.624749*h[65] + 1.196715, 0.624749*h[65] + 1.196715 > 0), (0.003124*h[65] + 0.005984, True))
h[67] = 0.135646
h[68] = 0.965753*h[24] + 0.050607
h[69] = -sin(0.90383*i[1] - 0.490802)
h[70] = 1.396789*h[46] + 0.015194
h[71] = -0.084691
h[72] = 1.0*h[14]
h[73] = 1.0*h[41]
h[74] = 1.0*i[0] + 0.076165
o[0] = -tanh(0.452189*i[0]*i[1] + 0.286043)
o[1] = tanh(0.594063*h[47] - 0.356644*i[0] + 0.224704)

In addition, when I load the best individual from a later generation, such as the best individual from generation 1500, a nan error occurs during LaTeX and Python code generation. However, the NEAT algorithm is still running well during iterations. Why does this issue occur when loading the best individual?
load best individual config:

import os
os.environ["CUDA_VISIBLE_DEVICES"] = "6"
import jax.numpy as jnp

from tensorneat.pipeline import Pipeline
from tensorneat.algorithm.neat import NEAT
from tensorneat.genome import DefaultGenome, BiasNode, DefaultMutation
from tensorneat.problem.func_fit import CustomFuncFit
from tensorneat.common import ACT, AGG
from tensorneat.problem.func_fit import custom2
import jax
import pickle

if __name__ == "__main__":
    pipeline = Pipeline(
        algorithm=NEAT(
            pop_size=1000,
            species_size=20,
            survival_threshold=0.025,
            min_species_size=2,
            genome=DefaultGenome(
                num_inputs=2,
                num_outputs=2,
                max_nodes=100,
                max_conns=250,
                # init_hidden_layers=(50,),20.1
                init_hidden_layers=(20,),
                mutation=DefaultMutation(
                    node_add=0.8,  # default is 0.1
                    node_delete=0.4,
                    conn_add=0.8,  # default is 0.2
                    conn_delete=0.4,
                ),
                node_gene=BiasNode(
                    activation_options=[ACT.lelu, ACT.tanh, ACT.sin, ACT.relu, ACT.identity],
                    aggregation_options=[AGG.sum, AGG.product, AGG.mean],
                ),
                output_transform=ACT.tanh,
            ),
        ),
        problem=custom2(),
        generation_limit=500,
        fitness_target=-1e-6,
        seed=45,
    )
    state = pipeline.setup()

    with open("Pipeline 20250317164510/genomes/1899.pkl", "rb") as f:
        best = pickle.load(f)

    best = jax.device_put(best)  # convert best to jax array on gpu
    transformed = pipeline.algorithm.genome.transform(state, *best)
    network = pipeline.algorithm.genome.network_dict(state, *best)
    print(pipeline.algorithm.genome.repr(state, *best))
    pipeline.algorithm.genome.visualize(network, save_path="./xor_network_44.svg")

    from tensorneat.common.sympy_tools import to_latex_code, to_python_code

    sympy_res = pipeline.algorithm.genome.sympy_func(
        state,
        network,
        sympy_output_transform=ACT.obtain_sympy(ACT.tanh)
    )

    print("\nLatex Formula:")
    latex_code = to_latex_code(*sympy_res)
    print(latex_code)

    print("\nPython Code:")
    python_code = to_python_code(*sympy_res)
    print(python_code)

error:

Traceback (most recent call last):
  File "/home/longchen/tensorneat/examples/func_fit2/visual.py", line 80, in <module>
    sympy_res = pipeline.algorithm.genome.sympy_func(
  File "/home/longchen/tensorneat/src/tensorneat/genome/default.py", line 206, in sympy_func
    nodes_exprs[symbols[i]], a_s = self.node_gene.sympy_func(
  File "/home/longchen/tensorneat/src/tensorneat/genome/gene/node/bias.py", line 181, in sympy_func
    z = ACT.obtain_sympy(node_dict["act"])(z)
  File "/home/longchen/miniconda3/envs/neat/lib/python3.10/site-packages/sympy/core/cache.py", line 72, in wrapper
    retval = cfunc(*args, **kwargs)
  File "/home/longchen/miniconda3/envs/neat/lib/python3.10/site-packages/sympy/core/function.py", line 466, in __new__
    result = super().__new__(cls, *args, **options)
  File "/home/longchen/miniconda3/envs/neat/lib/python3.10/site-packages/sympy/core/cache.py", line 72, in wrapper
    retval = cfunc(*args, **kwargs)
  File "/home/longchen/miniconda3/envs/neat/lib/python3.10/site-packages/sympy/core/function.py", line 307, in __new__
    evaluated = cls.eval(*args)
  File "/home/longchen/tensorneat/src/tensorneat/common/functions/act_sympy.py", line 47, in eval
    **return sp.Piecewise((z, z > 0), (leaky * z, True))**
  File "/home/longchen/miniconda3/envs/neat/lib/python3.10/site-packages/sympy/core/decorators.py", line 236, in _func
    return func(self, other)
  File "/home/longchen/miniconda3/envs/neat/lib/python3.10/site-packages/sympy/core/expr.py", line 360, in __gt__
    return StrictGreaterThan(self, other)
  File "/home/longchen/miniconda3/envs/neat/lib/python3.10/site-packages/sympy/core/relational.py", line 843, in __new__
    raise TypeError("Invalid NaN comparison")
TypeError: Invalid NaN comparison

[Polarisamoon]

Thank you for your reply. To facilitate loading historical best individuals, I modified the pipeline's analysis saving method. This should be correct?

def analysis(self, state, pop, fitnesses):
       .......
        if self.is_save:
            # save best
            best_genome = jax.device_get((pop[0][max_idx], pop[1][max_idx]))
            file_name = os.path.join(self.genome_dir, f"{generation}.pkl")
            # with open(file_name, "wb") as f:
            #     np.savez(
            #         f,
            #         nodes=best_genome[0],
            #         conns=best_genome[1],
            #         fitness=self.best_fitness,
            #     )
            # best = jax.device_get(best)  # convert best to numpy array on cpu
            with open(file_name, "wb") as f:
                pickle.dump(best_genome, f)
     ........

[WLS2002]

It is indeed a bug and I have already fixed it! Please re-install and try again. Thank you for your support to TensorNEAT!

[Polarisamoon]

Thank you very much, but I still have an issue — when generating the formulas for some individuals, an error occurs with NaN values.

Traceback (most recent call last):
  File "/home/longchen/tensorneat/examples/func_fit2/visual.py", line 80, in <module>
    sympy_res = pipeline.algorithm.genome.sympy_func(
  File "/home/longchen/tensorneat/src/tensorneat/genome/default.py", line 206, in sympy_func
    nodes_exprs[symbols[i]], a_s = self.node_gene.sympy_func(
  File "/home/longchen/tensorneat/src/tensorneat/genome/gene/node/bias.py", line 181, in sympy_func
    z = ACT.obtain_sympy(node_dict["act"])(z)
  File "/home/longchen/miniconda3/envs/neat/lib/python3.10/site-packages/sympy/core/cache.py", line 72, in wrapper
    retval = cfunc(*args, **kwargs)
  File "/home/longchen/miniconda3/envs/neat/lib/python3.10/site-packages/sympy/core/function.py", line 466, in __new__
    result = super().__new__(cls, *args, **options)
  File "/home/longchen/miniconda3/envs/neat/lib/python3.10/site-packages/sympy/core/cache.py", line 72, in wrapper
    retval = cfunc(*args, **kwargs)
  File "/home/longchen/miniconda3/envs/neat/lib/python3.10/site-packages/sympy/core/function.py", line 307, in __new__
    evaluated = cls.eval(*args)
  File "/home/longchen/tensorneat/src/tensorneat/common/functions/act_sympy.py", line 47, in eval
    **return sp.Piecewise((z, z > 0), (leaky * z, True))**
  File "/home/longchen/miniconda3/envs/neat/lib/python3.10/site-packages/sympy/core/decorators.py", line 236, in _func
    return func(self, other)
  File "/home/longchen/miniconda3/envs/neat/lib/python3.10/site-packages/sympy/core/expr.py", line 360, in __gt__
    return StrictGreaterThan(self, other)
  File "/home/longchen/miniconda3/envs/neat/lib/python3.10/site-packages/sympy/core/relational.py", line 843, in __new__
    raise TypeError("Invalid NaN comparison")
TypeError: Invalid NaN comparison

When I change the forward function of the faulty activation function to return x, I find that the issue is usually caused by the first few generated node values being nan.

Python Code:
h = np.zeros(83)
o = np.zeros(2)
h[5] = -0.052804
**h[38] = nan**
h[57] = 0.642575
**h[67] = nan**
h[72] = 0.513136
h[75] = 0.981454
o[0] = tanh(0.429377*i[0] + 0.52406*i[1] - 0.219799)
h[1] = -0.245492*i[1] - 0.559445
h[2] = 0.424013*i[0]*i[1] - 0.512945
h[3] = 0.171796*i[0] - 0.556728*i[1] + 0.3537
h[4] = -0.443455*i[0]*i[1] + 1.497627
h[8] = -1.360914*i[0] + 0.26604*i[1] + 0.191714
h[11] = sin(-2.091658*i[0] + 0.3995*i[1] + 0.600933)
h[12] = -0.396824*i[0] + 0.17315*i[1] + 1.303732
h[14] = 1.0*h[67]
h[21] = 0.435766*i[1] + 0.472511
h[27] = -0.712874*i[0] - 0.495742
h[28] = sin(1.460719*i[1] - 0.24362)
h[29] = -0.19752*i[0] - 0.254908*i[1] - 0.106494
h[31] = -tanh(0.141044*i[0] + 1.775136*i[1] + 0.36485)
h[32] = -0.454434*i[0] - 1.151417*i[1] + 0.674136
h[35] = -sin(0.288719*i[0] + 0.803896*i[1] - 1.166888)
h[36] = 0.926941*i[1] + 0.281859
h[40] = 0.023819*i[0] - 2.563115*i[1] + 1.459887
h[43] = 1.129956*i[0] + 1.867276*i[1] + 0.144179
h[44] = -4.191634*i[0] + 1.579335*i[1] - 0.383226
h[47] = -0.652353*h[75] + 2.009886*i[0] + 0.225796*i[1] - 0.060314
h[50] = 0.655507*i[1] - 0.189952
h[52] = -0.147537*h[72]*i[0] - 0.252661
h[53] = sin(1.03652*i[1] + 1.253294)
h[55] = sin(0.523939*i[1] + 0.013736)
h[56] = -tanh(0.647628*i[0] - 0.830922)
h[63] = -0.760447*i[1] - 1.948519
h[69] = 0.081852 - 1.720639*i[1]
h[70] = 1.0*h[38] - 0.09633
h[73] = 0.68243*i[1] - 0.053929
h[74] = -1.312601*i[1] - 0.050778
h[13] = 1.736833*h[73] + 3.475738*i[0] - 1.066896
h[15] = 2.794257*h[63] + 1.301596
h[22] = 1.041002*h[74] - 0.305383
h[25] = -sin(1.040145*h[55] + 1.844719)
h[30] = 1.815945*h[1]*i[0] + 1.418536
h[33] = 0.987256*h[53]*i[0] - 0.520329
h[34] = tanh(0.998788*h[53] + 0.735281*i[0] - 1.43406*i[1] + 0.510802)
h[39] = sin(0.195217*h[40]*i[0] + 0.080745)
h[41] = 0.662652*h[44] - 0.211234
h[42] = -1.61414*h[69] - 0.082936*h[72] + 1.610071*i[0] - 0.52406
h[46] = 0.086345*h[43] - 0.271526*h[57] + 0.244069*h[70] - 0.083041
h[51] = 1.626549*h[27] + 0.30748
h[58] = 1.071474*h[56] - 0.017977
h[64] = 1.306868*h[43] + 0.665303
h[71] = 0.965996*h[47] - 0.173041
h[79] = 1.00188*h[36]
h[80] = 0.462743 - 0.828579*h[31]
h[82] = 0.944536*h[3] - 0.282159
o[1] = tanh(0.45328*h[71] - 0.995943*i[0] + 0.058798)
h[26] = 0.860774*h[34] - 0.319913
h[37] = -tanh(0.304758*h[15] + 1.182739*i[0] + 0.685101)
h[54] = sin(1.520915*h[41] + 0.83524)
h[59] = 1.557329*h[39]*i[1] - 0.515575
h[62] = 0.398659*h[14] - 0.991241*h[58] - 0.866367
h[65] = 0.04128*h[42] + 0.547425*i[1] - 1.317257
h[76] = 1.278675*h[25] + 0.445822
h[78] = 0.21301*h[82] - 0.107387
h[81] = 0.924417*h[33] - 0.020344
h[0] = tanh(0.514568*h[59]*i[1] - 0.195821)
h[6] = -sin(1.294495*h[28] + 1.351636*h[42] - 0.464762*o[1] + 1.143229)
h[9] = 1.75279*h[21] + 0.640222*h[54] - 0.287071*h[56] + 0.687091*i[0] + 0.198053
h[19] = -sin(-1.563798*h[59] + 0.932974*i[0] + 0.073652*i[1] + 0.76119)
h[20] = 0.649488*h[22]*h[27]*i[0]*o[1] + 3.558526
h[23] = -0.197316*h[46]*h[54] + 0.418702
h[60] = 0.407264*h[59] - 0.004924
h[77] = 0.956375*h[26] + 0.324365
h[7] = -sin(0.027146*h[50]*h[60]*i[0] + 0.298587)
h[16] = 0.950208*h[9] - 0.308087
h[18] = 0.138253
h[61] = 0.860832 - 1.601575*h[60]
h[24] = 1.62276*h[37]*h[61]*i[0] - 0.16833
h[45] = tanh(1.66534*h[18]*i[0]*i[1] + 1.131848)
h[10] = 0.080952*h[45]*o[1] - 0.391998
h[17] = 0.806623*h[45] - 0.410568*i[0] + 2.312075*i[1] + 1.357916
h[48] = 0.058589*h[0] - 0.016585*h[12] + 0.109061*h[13] + 0.039001*h[2] - 0.033365*h[20] + 0.020796*h[23] - 0.022747*h[24] + 0.048316*h[25] - 0.036998*h[27] + 0.04974*h[29] + 0.043351*h[30] + 0.046492*h[32] + 0.029253*h[33] + 0.044255*h[35] + 0.032847*h[4] + 0.010767*h[40] - 0.034032*h[42] - 0.040208*h[43] - 0.144269*h[44] + 0.031318*h[45] + 0.032753*h[47] - 0.063265*h[6] - 0.009671*h[7] + 0.018715*h[78] + 0.012476*h[79] - 0.020222*h[8] - 0.013831*h[80] + 0.069409*h[9] - 0.004067*o[0] - 0.016491
h[68] = -0.448209
h[49] = 0.00237*h[0] - 0.001594*h[10] + 0.010374*h[11] + 0.026331*h[12] + 0.071027*h[13] + 0.165107*h[16] + 0.0786*h[17] - 0.084129*h[19] - 0.107403*h[29] - 0.016015*h[3] - 0.024895*h[30] - 0.072011*h[35] + 0.040803*h[36] + 0.038442*h[37] + 0.062023*h[4] - 0.100208*h[40] - 0.015513*h[42] + 0.00612*h[47] - 0.069744*h[51] + 0.000773*h[54] + 0.023279*h[6] + 0.080606*h[64] + 0.045874*h[7] + 0.019967*h[72] + 0.037247*h[76] + 0.059564*h[8] - 0.044619*h[81] + 0.033209*o[0] - 0.019919*o[1] - 0.114961
h[66] = 0.107958*h[3]*h[48]*h[65] + 0.059296

[WLS2002]

Hi, sorry for the late reply. I was busy recently and just got a chance to fix the issue.

Thank you for your patience, and thanks again for reporting it. Let me know if anything else comes up.