pyMOTO 1.3.0__py3-none-any.whl → 1.4.0__py3-none-any.whl

pymoto/core_objects.py

@@ -1,9 +1,11 @@
-from typing import Union, List, Any
+import sys
 import warnings
 import inspect
 import time
-from .utils import _parse_to_list, _concatenate_to_array, _split_from_array
+import copy
+from typing import Union, List, Any
 from abc import ABC, abstractmethod
+from .utils import _parse_to_list, _concatenate_to_array, _split_from_array
 
 
 # Local helper functions
@@ -11,7 +13,7 @@ def err_fmt(*args):
     """ Format error strings for locating Modules and Signals"""
     err_str = ""
     for a in args:
-        err_str += f"\n\t[ {a} ]"
+        err_str += f"\n\t| {a}"
     return err_str
 
 
@@ -76,16 +78,20 @@ class Signal:
     >> Signal(tag='x2')
 
     """
-    def __init__(self, tag: str = "", state: Any = None, sensitivity: Any = None):
+    def __init__(self, tag: str = "", state: Any = None, sensitivity: Any = None, min: Any = None, max: Any = None):
         """
-
-        :param tag: The name of the signal (string)
-        :param state: The initialized state (optional)
-        :param sensitivity: The initialized sensitivity (optional)
+        Keyword Args:
+            tag: The name of the signal
+            state: The initialized state
+            sensitivity: The initialized sensitivity
+            min: Minimum allowed value
+            max: Maximum allowed value
         """
         self.tag = tag
         self.state = state
         self.sensitivity = sensitivity
+        self.min = min
+        self.max = max
         self.keep_alloc = sensitivity is not None
 
         # Save error string to location where it is initialized
@@ -95,11 +101,12 @@ class Signal:
         return err_fmt(f"Signal \'{self.tag}\', initialized in {self._init_loc}")
 
     def add_sensitivity(self, ds: Any):
+        """ Add a new term to internal sensitivity """
         try:
             if ds is None:
                 return
             if self.sensitivity is None:
-                self.sensitivity = ds
+                self.sensitivity = copy.deepcopy(ds)
             else:
                 self.sensitivity += ds
             return self
@@ -116,8 +123,12 @@ class Signal:
     def reset(self, keep_alloc: bool = None):
         """ Reset the sensitivities to zero or None
         This must be called to clear internal memory of subsequent sensitivity calculations.
-        :param keep_alloc: Keep the sensitivity allocation intact?
-        :return: self
+
+        Args:
+            keep_alloc: Keep the sensitivity allocation intact?
+
+        Returns:
+            self
         """
         if self.sensitivity is None:
             return self
@@ -138,11 +149,34 @@ class Signal:
 
     def __getitem__(self, item):
         """ Obtain a sliced signal, for using its partial contents.
-        :param item: Slice indices
-        :return: Sliced signal (SignalSlice)
+
+        Args:
+            item: Slice indices
+
+        Returns:
+            Sliced signal (SignalSlice)
         """
         return SignalSlice(self, item)
 
+    def __str__(self):
+        state_msg = f"state {self.state}" if self.state is not None else "empty state"
+        state_msg = state_msg.split('\n')
+        if len(state_msg) > 1:
+            state_msg = state_msg[0] + ' ... ' + state_msg[-1]
+        else:
+            state_msg = state_msg[0]
+        return f"Signal \"{self.tag}\" with {state_msg}"
+
+    def __repr__(self):
+        state_msg = f"state {self.state}" if self.state is not None else "empty state"
+        state_msg = state_msg.split('\n')
+        if len(state_msg) > 1:
+            state_msg = state_msg[0] + ' ... ' + state_msg[-1]
+        else:
+            state_msg = state_msg[0]
+        sens_msg = 'empty sensitivity' if self.sensitivity is None else 'non-empty sensitivity'
+        return f"Signal \"{self.tag}\" with {state_msg} and {sens_msg} at {hex(id(self))}"
+
 
 class SignalSlice(Signal):
     """ Slice operator for a Signal
@@ -169,7 +203,8 @@ class SignalSlice(Signal):
             return None if self.orig_signal.state is None else self.orig_signal.state[self.slice]
         except Exception as e:
             # Possibilities: Unslicable object (TypeError) or Wrong dimensions or out of range (IndexError)
-            raise type(e)("SignalSlice.state (getter)" + self._err_str()) from e
+            raise type(e)(str(e) + "\n\t| Above error was raised in SignalSlice.state (getter). Signal details:" +
+                      self._err_str()).with_traceback(sys.exc_info()[2])
 
     @state.setter
     def state(self, new_state):
@@ -177,7 +212,8 @@ class SignalSlice(Signal):
             self.orig_signal.state[self.slice] = new_state
         except Exception as e:
             # Possibilities: Unslicable object (TypeError) or Wrong dimensions or out of range (IndexError)
-            raise type(e)("SignalSlice.state (setter)" + self._err_str()) from e
+            raise type(e)(str(e) + "\n\t| Above error was raised in SignalSlice.state (setter). Signal details:" +
+                          self._err_str()).with_traceback(sys.exc_info()[2])
 
     @property
     def sensitivity(self):
@@ -185,7 +221,8 @@ class SignalSlice(Signal):
             return None if self.orig_signal.sensitivity is None else self.orig_signal.sensitivity[self.slice]
         except Exception as e:
             # Possibilities: Unslicable object (TypeError) or Wrong dimensions or out of range (IndexError)
-            raise type(e)("SignalSlice.sensitivity (getter)" + self._err_str()) from e
+            raise type(e)(str(e) + "\n\t| Above error was raised in SignalSlice.sensitivity (getter). Signal details:" +
+                          self._err_str()).with_traceback(sys.exc_info()[2])
 
     @sensitivity.setter
     def sensitivity(self, new_sens):
@@ -207,7 +244,8 @@ class SignalSlice(Signal):
             self.orig_signal.sensitivity[self.slice] = new_sens
         except Exception as e:
             # Possibilities: Unslicable object (TypeError) or Wrong dimensions or out of range (IndexError)
-            raise type(e)("SignalSlice.sensitivity (setter)" + self._err_str()) from e
+            raise type(e)(str(e) + "\n\t| Above error was raised in SignalSlice.state (setter). Signal details:" +
+                          self._err_str()).with_traceback(sys.exc_info()[2])
 
     def reset(self, keep_alloc: bool = None):
         """ Reset the sensitivities to zero or None
@@ -231,7 +269,7 @@ def make_signals(*args):
     return ret
 
 
-def _check_valid_signal(sig: Any):
+def _is_valid_signal(sig: Any):
     """ Checks if the argument is a valid Signal object
     :param sig: The object to check
     :return: True if it is a valid Signal
@@ -240,10 +278,10 @@ def _check_valid_signal(sig: Any):
         return True
     if all([hasattr(sig, f) for f in ["state", "sensitivity", "add_sensitivity", "reset"]]):
         return True
-    raise TypeError(f"Given argument with type \'{type(sig).__name__}\' is not a valid Signal")
+    return False
 
 
-def _check_valid_module(mod: Any):
+def _is_valid_module(mod: Any):
     """ Checks if the argument is a valid Module object
     :param mod: The object to check
     :return: True if it is a valid Module
@@ -252,7 +290,7 @@ def _check_valid_module(mod: Any):
         return True
     if hasattr(mod, "response") and hasattr(mod, "sensitivity") and hasattr(mod, "reset"):
         return True
-    raise TypeError(f"Given argument with type \'{type(mod).__name__}\' is not a valid Module")
+    return False
 
 
 def _check_function_signature(fn, signals):
@@ -370,60 +408,48 @@ class Module(ABC, RegisteredClass):
     >> Module(sig_in=[inputs], sig_out=[outputs]
     """
 
-    def _err_str(self, init: bool = True, add_signal: bool = True, fn=None):
+    def _err_str(self, module_signature: bool = True, init: bool = True, fn=None):
         str_list = []
-        if init:
-            str_list.append(f"Module \'{type(self).__name__}\', initialized in {self._init_loc}")
-        if add_signal:
+
+        if module_signature:
             inp_str = "Inputs: " + ", ".join([s.tag if hasattr(s, 'tag') else 'N/A' for s in self.sig_in]) if len(self.sig_in) > 0 else "No inputs"
             out_str = "Outputs: " + ", ".join([s.tag if hasattr(s, 'tag') else 'N/A' for s in self.sig_out]) if len(self.sig_out) > 0 else "No outputs"
-            str_list.append(inp_str + " --> " + out_str)
+            str_list.append(f"Module \'{type(self).__name__}\'( " + inp_str + " ) --> " + out_str)
+        if init:
+            str_list.append(f"Used in {self._init_loc}")
         if fn is not None:
             name = f"{fn.__self__.__class__.__name__}.{fn.__name__}{inspect.signature(fn)}"
             lineno = inspect.getsourcelines(fn)[1]
             filename = inspect.getfile(fn)
-            str_list.append(f"Implemented in File \"{filename}\", line {lineno}, in {name}")
+            str_list.append(f"Implementation in File \"{filename}\", line {lineno}, in {name}")
         return err_fmt(*str_list)
 
     # flake8: noqa: C901
     def __init__(self, sig_in: Union[Signal, List[Signal]] = None, sig_out: Union[Signal, List[Signal]] = None,
                  *args, **kwargs):
-        # TODO: Reduce complexity of this init
         self._init_loc = get_init_str()
 
         self.sig_in = _parse_to_list(sig_in)
         self.sig_out = _parse_to_list(sig_out)
         for i, s in enumerate(self.sig_in):
-            try:
-                _check_valid_signal(s)
-            except Exception as e:
-                earg0 = e.args[0] if len(e.args) > 0 else ''
-                earg1 = e.args[1:] if len(e.args) > 1 else ()
-                raise type(e)(f"Invalid input signal #{i+1} - " + str(earg0) + self._err_str(), *earg1) from None
+            if not _is_valid_signal(s):
+                tag = f" (\'{s.tag}\')" if hasattr(s, 'tag') else ''
+                raise TypeError(f"Input {i}{tag} is not a valid signal, type=\'{type(s).__name__}\'.")
 
         for i, s in enumerate(self.sig_out):
-            try:
-                _check_valid_signal(s)
-            except Exception as e:
-                earg0 = e.args[0] if len(e.args) > 0 else ''
-                earg1 = e.args[1:] if len(e.args) > 1 else ()
-                raise type(e)(f"Invalid output signal #{i+1} - " + str(earg0) + self._err_str(), *earg1) from None
+            if not _is_valid_signal(s):
+                tag = f" (\'{s.tag}\')" if hasattr(s, 'tag') else ''
+                raise TypeError(f"Output {i}{tag} is not a valid signal, type=\'{type(s).__name__}\'.")
 
-        try:
-            # Call preparation of submodule with remaining arguments
-            self._prepare(*args, **kwargs)
-        except Exception as e:
-            earg0 = e.args[0] if len(e.args) > 0 else ''
-            earg1 = e.args[1:] if len(e.args) > 1 else ()
-            raise type(e)("_prepare() - " + str(earg0) + self._err_str(fn=self._prepare), *earg1) from e
+        # Call preparation of submodule with remaining arguments
+        self._prepare(*args, **kwargs)
 
         try:
             # Check if the signals match _response() signature
             _check_function_signature(self._response, self.sig_in)
         except Exception as e:
-            earg0 = e.args[0] if len(e.args) > 0 else ''
-            earg1 = e.args[1:] if len(e.args) > 1 else ()
-            raise type(e)(str(earg0) + self._err_str(fn=self._response), *earg1) from None
+            raise type(e)(str(e) + "\n\t| Module details:" +
+                          self._err_str(fn=self._response)).with_traceback(sys.exc_info()[2])
 
         try:
             # If no output signals are given, but are required, try to initialize them here
@@ -441,9 +467,8 @@ class Module(ABC, RegisteredClass):
             # Check if signals match _sensitivity() signature
             _check_function_signature(self._sensitivity, self.sig_out)
         except Exception as e:
-            earg0 = e.args[0] if len(e.args) > 0 else ''
-            earg1 = e.args[1:] if len(e.args) > 1 else ()
-            raise type(e)(str(earg0) + self._err_str(fn=self._sensitivity), *earg1) from None
+            raise type(e)(str(e) + "\n\t| Module details:" +
+                          self._err_str(fn=self._sensitivity)).with_traceback(sys.exc_info()[2])
 
     def response(self):
         """ Calculate the response from sig_in and output this to sig_out """
@@ -461,9 +486,9 @@ class Module(ABC, RegisteredClass):
                 self.sig_out[i].state = val
             return self
         except Exception as e:
-            earg0 = e.args[0] if len(e.args) > 0 else ''
-            earg1 = e.args[1:] if len(e.args) > 1 else ()
-            raise type(e)("response() - " + str(earg0) + self._err_str(fn=self._response), *earg1) from e
+            # https://stackoverflow.com/questions/6062576/adding-information-to-an-exception
+            raise type(e)(str(e) + "\n\t| Above error was raised when calling response(). Module details:" +
+                          self._err_str(fn=self._response)).with_traceback(sys.exc_info()[2])
 
     def __call__(self):
         return self.response()
@@ -494,9 +519,8 @@ class Module(ABC, RegisteredClass):
 
             return self
         except Exception as e:
-            earg0 = e.args[0] if len(e.args) > 0 else ''
-            earg1 = e.args[1:] if len(e.args) > 1 else ()
-            raise type(e)("sensitivity() - " + str(earg0) + self._err_str(fn=self._sensitivity), *earg1) from e
+            raise type(e)(str(e) + "\n\t| Above error was raised when calling sensitivity(). Module details:" +
+                          self._err_str(fn=self._sensitivity)).with_traceback(sys.exc_info()[2])
 
     def reset(self):
         """ Reset the state of the sensitivities (they are set to zero or to None) """
@@ -506,9 +530,8 @@ class Module(ABC, RegisteredClass):
             self._reset()
             return self
         except Exception as e:
-            earg0 = e.args[0] if len(e.args) > 0 else ''
-            earg1 = e.args[1:] if len(e.args) > 1 else ()
-            raise type(e)("reset() - " + str(earg0) + self._err_str(fn=self._reset), *earg1) from e
+            raise type(e)(str(e) + "\n\t| Above error was raised when calling reset(). Module details:" +
+                          self._err_str(fn=self._response)).with_traceback(sys.exc_info()[2])
 
     # METHODS TO BE DEFINED BY USER
     def _prepare(self, *args, **kwargs):
@@ -542,35 +565,33 @@ class Network(Module):
     """
     def __init__(self, *args, print_timing=False):
         self._init_loc = get_init_str()
-        try:
-            # Obtain the internal blocks
-            self.mods = _parse_to_list(*args)
-
-            # Check if the blocks are initialized, else create them
-            for i, b in enumerate(self.mods):
-                if isinstance(b, dict):
-                    exclude_keys = ['type']
-                    b_ex = {k: b[k] for k in set(list(b.keys())) - set(exclude_keys)}
-                    self.mods[i] = Module.create(b['type'], **b_ex)
-
-            # Check validity of modules
-            [_check_valid_module(m) for m in self.mods]
-
-            # Gather all the input and output signals of the internal blocks
-            all_in = set()
-            all_out = set()
-            [all_in.update(b.sig_in) for b in self.mods]
-            [all_out.update(b.sig_out) for b in self.mods]
-            in_unique = all_in - all_out
-
-            # Initialize the parent module, with correct inputs and outputs
-            super().__init__(list(in_unique), list(all_out))
-
-            self.print_timing = print_timing
-        except Exception as e:
-            earg0 = e.args[0] if len(e.args) > 0 else ''
-            earg1 = e.args[1:] if len(e.args) > 1 else ()
-            raise type(e)(str(earg0) + self._err_str(add_signal=False), *earg1) from None
+
+        # Obtain the internal blocks
+        self.mods = _parse_to_list(*args)
+
+        # Check if the blocks are initialized, else create them
+        for i, b in enumerate(self.mods):
+            if isinstance(b, dict):
+                exclude_keys = ['type']
+                b_ex = {k: b[k] for k in set(list(b.keys())) - set(exclude_keys)}
+                self.mods[i] = Module.create(b['type'], **b_ex)
+
+        # Check validity of modules
+        for m in self.mods:
+            if not _is_valid_module(m):
+                raise TypeError(f"Argument is not a valid Module, type=\'{type(mod).__name__}\'.")
+
+        # Gather all the input and output signals of the internal blocks
+        all_in = set()
+        all_out = set()
+        [all_in.update(b.sig_in) for b in self.mods]
+        [all_out.update(b.sig_out) for b in self.mods]
+        in_unique = all_in - all_out
+
+        # Initialize the parent module, with correct inputs and outputs
+        super().__init__(list(in_unique), list(all_out))
+
+        self.print_timing = print_timing
 
     def timefn(self, fn):
         start_t = time.time()
@@ -611,12 +632,9 @@ class Network(Module):
         modlist = _parse_to_list(*newmods)
 
         # Check if the blocks are initialized, else create them
-        for i, b in enumerate(modlist):
-            try:  # Check validity of modules
-                _check_valid_module(b)
-            except Exception as e:
-                raise type(e)("append() - Trying to append invalid module " + str(e.args[0])
-                              + self._err_str(add_signal=False), *e.args[1:]) from None
+        for i, m in enumerate(modlist):
+            if not _is_valid_module(m):
+                raise TypeError(f"Argument #{i} is not a valid module, type=\'{type(mod).__name__}\'.")
 
         # Obtain the internal blocks
         self.mods.extend(modlist)
@@ -624,25 +642,11 @@ class Network(Module):
         # Gather all the input and output signals of the internal blocks
         all_in = set()
         all_out = set()
-        [all_in.update(b.sig_in) for b in self.mods]
-        [all_out.update(b.sig_out) for b in self.mods]
+        [all_in.update(m.sig_in) for m in self.mods]
+        [all_out.update(m.sig_out) for m in self.mods]
         in_unique = all_in - all_out
 
         self.sig_in = _parse_to_list(in_unique)
-        try:
-            [_check_valid_signal(s) for s in self.sig_in]
-        except Exception as e:
-            earg0 = e.args[0] if len(e.args) > 0 else ''
-            earg1 = e.args[1:] if len(e.args) > 1 else ()
-            raise type(e)("append() - Invalid input signals " + str(earg0)
-                          + self._err_str(add_signal=False), *earg1) from None
         self.sig_out = _parse_to_list(all_out)
-        try:
-            [_check_valid_signal(s) for s in self.sig_out]
-        except Exception as e:
-            earg0 = e.args[0] if len(e.args) > 0 else ''
-            earg1 = e.args[1:] if len(e.args) > 1 else ()
-            raise type(e)("append() - Invalid output signals " + str(earg0)
-                          + self._err_str(add_signal=False), *earg1) from None
 
         return modlist[-1].sig_out[0] if len(modlist[-1].sig_out) == 1 else modlist[-1].sig_out  # Returns the output signal

pymoto/modules/aggregation.py

@@ -0,0 +1,209 @@
+import warnings
+import abc
+import numpy as np
+import scipy.special as spsp
+from pymoto import Module
+
+
+class AggActiveSet:
+    """ Determine active set by discarding lower or upper fraction of a set of values
+
+    Args:
+       lower_rel: Fraction of values closest to minimum to discard (based on value)
+       upper_rel: Fraction of values closest to maximum to discard (based on value)
+       lower_amt: Fraction of lowest values to discard (based on sorting)
+       upper_amt: Fraction of highest values to discard (based on sorting)
+    """
+    def __init__(self, lower_rel=0.0, upper_rel=1.0, lower_amt=0.0, upper_amt=1.0):
+        assert upper_rel > lower_rel, "Upper must be larger than lower to keep values in the set"
+        assert upper_amt > lower_amt, "Upper must be larger than lower to keep values in the set"
+        self.lower_rel, self.upper_rel = lower_rel, upper_rel
+        self.lower_amt, self.upper_amt = lower_amt, upper_amt
+
+    def __call__(self, x):
+        """ Generate an active set for given array """
+        xmin, xmax = np.min(x), np.max(x)
+        if (xmax - xmin) == 0:  # All values are the same, so no active set can be taken
+            return Ellipsis
+
+        sel = np.ones_like(x, dtype=bool)
+
+        # Select based on value
+        xrel = (x - xmin) / (xmax - xmin)  # Normalize between 0 and 1
+        if self.lower_rel > 0:
+            sel = np.logical_and(sel, xrel >= self.lower_rel)
+        if self.upper_rel < 1:
+            sel = np.logical_and(sel, xrel <= self.upper_rel)
+
+        # Remove lowest and highest N values
+        i_sort = np.argsort(x)
+        if self.lower_amt > 0:
+            n_lower_amt = int(x.size * self.lower_amt)
+            sel[i_sort[:n_lower_amt]] = False
+
+        if self.upper_amt < 1:
+            n_upper_amt = int(x.size * (1 - self.upper_amt))
+            sel[i_sort[-n_upper_amt:]] = False
+
+        return sel
+
+
+class AggScaling:
+    """ Scaling strategy to absolute minimum or maximum
+
+    Args:
+        which: Scale to `min` or `max`
+        damping(optional): Damping factor between [0, 1), for a value of 0.0 the aggregation approximation is corrected
+          to the exact maximum or minimum of the input set
+    """
+    def __init__(self, which: str, damping=0.0):
+        self.damping = damping
+        if which.lower() == 'min':
+            self.f = np.min
+        elif which.lower() == 'max':
+            self.f = np.max
+        else:
+            raise ValueError("Argument `which` can only be 'min' or 'max'")
+        self.sf = None
+
+    def __call__(self, x, fx_approx):
+        """ Determine scaling factor
+
+        Args:
+            x: Set of values
+            fx_approx: Approximated minimum / maximum
+
+        Returns:
+            Scaling factor
+        """
+        trueval = self.f(x)
+        scale = trueval / fx_approx
+        if self.sf is None:
+            self.sf = scale
+        else:
+            self.sf = self.damping * self.sf + (1 - self.damping) * scale
+        return self.sf
+
+
+class Aggregation(Module):
+    """ Generic Aggregation module (cannot be used directly, but can only be used as superclass)
+
+    Keyword Args:
+        scaling(optional): Scaling strategy to improve approximation :py:class:`pymoto.AggScaling`
+        active_set(optional): Active set strategy to improve approximation :py:class:`pymoto.AggActiveSet`
+    """
+    def _prepare(self, scaling: AggScaling = None, active_set: AggActiveSet = None):
+        # This prepare function MUST be called in the _prepare function of sub-classes
+        self.scaling = scaling
+        self.active_set = active_set
+        self.sf = 1.0
+
+    @abc.abstractmethod
+    def aggregation_function(self, x):
+        """ Calculates f(x) """
+        raise NotImplementedError()
+
+    @abc.abstractmethod
+    def aggregation_derivative(self, x):
+        """" Calculates df(x) / dx """
+        raise NotImplementedError()
+
+    def _response(self, x):
+        # Determine active set
+        if self.active_set is not None:
+            self.select = self.active_set(x)
+        else:
+            self.select = Ellipsis
+
+        # Get aggregated value
+        xagg = self.aggregation_function(x[self.select])
+
+        # Scale
+        if self.scaling is not None:
+            self.sf = self.scaling(x[self.select], xagg)
+        return self.sf * xagg
+
+    def _sensitivity(self, dfdy):
+        x = self.sig_in[0].state
+        dydx = self.aggregation_derivative(x[self.select])
+        dx = np.zeros_like(x)
+        dx[self.select] += self.sf * dfdy * dydx
+        return dx
+
+
+class PNorm(Aggregation):
+    r""" P-norm aggregration
+
+    :math:`S_p(x_1, x_2, \dotsc, x_n) = \left( \sum_i (|x_i|^p) \right)^{1/p}
+
+    Only valid for positive :math:`x_i` when approximating the minimum or maximum
+
+    Args:
+        p: Power of the p-norm. Approximate maximum for `p>0` and minimum for `p<0`
+        scaling(optional): Scaling strategy to improve approximation :py:class:`pymoto.AggScaling`
+        active_set(optional): Active set strategy to improve approximation :py:class:`pymoto.AggActiveSet`
+    """
+    def _prepare(self, p=2, scaling: AggScaling = None, active_set: AggActiveSet = None):
+        self.p = p
+        self.y = None
+        super()._prepare(scaling, active_set)
+
+    def aggregation_function(self, x):
+        if np.min(x) < 0:
+            warnings.warn("PNorm is only valid for positive x")
+
+        # Get p-norm
+        return np.sum(np.abs(x) ** self.p) ** (1/self.p)
+
+    def aggregation_derivative(self, x):
+        pval = np.sum(np.abs(x) ** self.p) ** (1 / self.p - 1)
+        return pval * np.sign(x) * np.abs(x)**(self.p - 1)
+
+
+class SoftMinMax(Aggregation):
+    r""" Soft maximum/minimum function
+
+    :math:`S_a(x_1, x_2, \dotsc, x_n) = \frac{\sum_i (x_i \exp(a x_i))}{\sum_i (\exp(a x_i))}`
+
+    When using as maximum, it underestimates the maximum
+    It is exact however when :math:`x_1=x_2=\dotsc=x_n`
+
+    Args:
+        alpha: Scaling factor of the soft function. Approximate maximum for `alpha>0` and minimum for `alpha<0`
+        scaling(optional): Scaling strategy to improve approximation :py:class:`pymoto.AggScaling`
+        active_set(optional): Active set strategy to improve approximation :py:class:`pymoto.AggActiveSet`
+    """
+    def _prepare(self, alpha=1.0, scaling: AggScaling = None, active_set: AggActiveSet = None):
+        self.alpha = alpha
+        self.y = None
+        super()._prepare(scaling, active_set)
+
+    def aggregation_function(self, x):
+        self.y = np.sum(x * spsp.softmax(self.alpha * x))
+        return self.y
+
+    def aggregation_derivative(self, x):
+        return spsp.softmax(self.alpha * x) * (1 + self.alpha * (x - self.y))
+
+
+class KSFunction(Aggregation):
+    r""" Kreisselmeier and Steinhauser function from 1979
+
+    :math:`S_\rho(x_1, x_2, \dotsc, x_n) = \frac{1}{\rho} \ln \left( \sum_i \exp(\rho x_i) \right)`
+
+    Args:
+        rho: Scaling factor of the KS function. Approximate maximum for `rho>0` and minimum for `rho<0`
+        scaling(optional): Scaling strategy to improve approximation :py:class:`pymoto.AggScaling`
+        active_set(optional): Active set strategy to improve approximation :py:class:`pymoto.AggActiveSet`
+    """
+    def _prepare(self, rho=1.0, scaling: AggScaling = None, active_set: AggActiveSet = None):
+        self.rho = rho
+        self.y = None
+        super()._prepare(scaling, active_set)
+
+    def aggregation_function(self, x):
+        return 1/self.rho * np.log(np.sum(np.exp(self.rho * x)))
+
+    def aggregation_derivative(self, x):
+        erx = np.exp(self.rho * x)
+        return erx / np.sum(erx)