SymbolicRegression

Let give some symbols like x1 and x2. The algorithm will consider first the linear combinations a*x1+b and c*x2+d, then extra start symbolic expressions defined with extra_start_sym_expr. Set verbose to True to get some information during regression. During process, we can check some expressions with checked_sym_expr. It is possible to specify the cost function to optimize with elementwise_loss. auto_ops is used to select best operators among available operators, auto_ops_depth is a 2-tuple with depth for unary and binary operators. Then the first iteration starts:

• We apply the unary operators defined with unary_operators, for example e*log(a*x1+b)+f and g*log(c*x2+d)+h.
• We apply then the binary operators defined with binary_operators, for example i*(a*x1+b+c*x2+d)+j same as k*x1+l*x2+m, etc. We can skip some combinations according to symmetric binary operators defined with symmetric_binary_operators.

discard_previous_expr will discard previous expressions at each level of iteration, avoiding combinatory explosion. To avoid too this explosion, maxloss, maxexpr, group_expr_size and maxtask are provided. Use process_sym_expr to process only target expressions. We can define the depth of operators with operator_depth and the maximal complexity of searched expressions with maxcomplexity. We can apply some weights to operators for depth with op_weights. At each combination, we compute an optimal expression. If discrete_param_values is empty, we search optimal parameters with scipy.fit_curve (maxfev is used), else we use brute force algorithm if possible (brute_force_limit is used), else a smallest brute force problem, else a curve fit algorithm followed by genetic algorithm with deap. discrete_param_values accept integer or float values for each parameter, str to define integer ranges like "(3, 5)" for range(3, 5), or "(3, 6, 2)" for range(3, 6, 2), or real ranges like "[3, 5, 1]" for np.linspace(3, 5, 1). We process like that until the computed loss is less than epsloss. eps is used to round numeric values and compare with other expressions for example. We can subsitute some target expressions with subs_expr. We can also avoid some expressions with avoided_expr. It is possible to call user callback during process with callback.

To use multiprocess functionalities, functions must be pickable and freeze_support must be enabled. It is possible to export results to CSV file with csv_filename.

For more information, see existing tests. pytest -v If pytest-xdist and pytest-timeout are installed: pytest -v -n auto --timeout 1200