evolutionary-circuits

Automatically generates analog circuits using evolutionary algorithms.

Requires ngspice, python and pypy (Yes, both of them). Also probably only works on a linux OS.

Interesting stuff happens in cgp.py, circuits.py is used to communicate with ngspice, plotting.py is used to implement plotting with matplotlib when running with pypy. getch.py gets a single character from output.

Work is currently in progress and documentation is missing. This is currently hard to use and requires a knowledge of how to use SPICE.

Simulation can be stopped with a keyboard interrupt. Progress is saved after every generation and can be continued by just running the script again.

#Installing

##ngspice

Download ngpspice either the git version or the current stable release, whichever you want. Git release often has performance improvements and possibly new features, but it also comes with the best bugs. If you are new to using SPICE it might be better to install latest stable version(As of 2012-09-10 this is version 24).

Compile and install ngspice with following commands:

 $ mkdir release
 $ cd release
 $ ../configure --enable-xspice --disable-debug
 $ make
 $ sudo make install

#Usage

The program is used by saving simulation settings in separate file and running it with command: "pypy cgp.py ". See examples folder for example scripts.

Using pypy is necessary. CPython raises PicklingError, because it can't pickle functions inside classes, which is required for multithreading.

##Simulation settings

Every name starting with underscore("_") is ignored by cgp.py and they can be used for internal functions in settings file.

Required settings are:

• title: Title of the simulation, this will be the name of the folder where output is placed.
• max_parts: Maximum number of devices allowed.
• spice_commands: SPICE simulation commands, can be string if there is only one simulation or list of string for more than one simulation.
• parts: A dictionary of available parts.
• fitness_function: The goal function, that circuits are ranked against.

Optional settings are:

• common: Common constructs in SPICE simulation, eg. input resistance, output load...
• models: String of SPICE models.
• constraints: One function or a list of functions for constraints. Some entries in list can be None. Length of list must equal the number of SPICE simulations.
• population: Number of circuits in one generation. Default is 1000.
• nodes: Number of nodes where parts can be attached. Default is same as max_parts.
• elitism: Number of circuits copied straight into a new generation. Default is 1.
• mutation_rate: Probability of mutations. Default is 0.7.
• crossover_rate: Probability of crossovers. Default is 0.2.
• fitness_weight: Weightings for fitness functions. Can be either numbers or functions of type "lambda x,**kwargs", where x is simulation x-axis value, kwargs has keys: "extra_value", "generation", possibly more as they are added. List of dictionaries. One list elements corresponds to one SPICE simulation. Dictionary keys are measurements('v(n2)','i(n1')...), and values are the weights the measurement is multiplied. Value can also be a function that takes input value and returns a number.
• extra_value: List of tuples that are minimum and maximum of extra values that chromosome can hold. This is returned to fitness function as argument extra_value. This can be used as example for transition voltage of an inverter.
• log_file: Filename where simulation log is saved.
• plot_titles: List of dictionaries of plot titles.
• plot_yrange: List of dictionaries of plot Y-axis ranges. Can be useful if you turn the output plots into an animation, this avoids the rescaling of the axes.
• selection_weight: Higher values make better performing circuits being picked more often and smaller values make selection more fair, 0.0 makes selection completely random. Default is 1.0 and values should be bigger than zero. Usable range us probably from 0.5 to 1.5. Don't set this too high or entropy in the gene pool is lost and genetic algorithm doesn't converge. You should only touch this if you know what you are doing.
• constraint_weight: Weights for constraints, similar to fitness_weight. Score added is percentage of values not passing times this score.
• max_mutations: Maximum number of mutations without re-evaluating the fitness function.
• constraint_free_generations: Number of generations before constraint functions are enabled.
• gradual_constraints: Apply constraint function scores gradually, default is True.
• constraint_ramp: Number of generations before constraint scores are at maximum levels. Does nothing if gradual_constraints if False.
• random_circuits: Percentage of random circuits in new generation. Default is 0.01(1%).
• plot_every_generation: True to plot every generation even if the best score is same or worse than the last generation. Useful is weights increase when current generation increases, doesn't need to be enabled for "gradual_constraints" to work properly. Default is False.
• default_scoring: Use default scoring(Squared error difference and weighting).
• custom_scoring: User defined scoring function. None if none is given. Both user defined and default scoring can be active at the same time, in this case scores are summed. Function should take two arguments: "result" and **kwargs. result is dictionary of measurements, each dictionary value is tuple of two lists: simulation x-axis and y-axis values. Example: {'v(n2)':[0.0,0.1,0.2],[1.0,1.1,1.2]}. Should return a score that has a type of float.
• timeout: SPICE simulation default timeout. This is raised automatically if default is too low.
• seed: Seed circuits to be added in the first generation. All of the devices need to be in the parts dictionary. Can be a list or string.
• seed_copies: Copies of seed circuits.

First you need to decide what components you want to allow and add them to the "parts" dictionary. The dictionary key is the name of component in SPICE, and the value should be an another dictionary with contents:

'nodes': Number of leads in the component
'value': True/False, component has a value(Example: True for resistors and
capacitors, their values are resistance and capacitance.
False for example transistors)
'min': Exponent of the components minimum value. Minimum value will be
10^(min)
'max': Exponent of the components maximum value.
'spice': Extra options for spice. This is appended to the end of components
description in spice. This is used for example transistors model.

Next you need SPICE simulation commands and a optional part models. Ground node is hard coded in SPICE to be 0.

Add "print "(eg. "print v(n2)" or "print i(n2)") in the spice commands to get an output to the program. You can print more than one measurement from one spice simulation.

Other settings you should specify are the fitness function or a list of them if you have more than one simulation. These are the goals of the evolution, that circuits are scored against.

Definition of the fitness function is:

def fitness\_function(a,b,**kwargs)

First argument is the input value of the measurement(eg. voltage, current, decibels, phase...). Second is the name of the SPICE measurement(eg. v(n2), vdb(n3), i(vin)). **kwargs has the optional extra_value of the chromosome and the current generation.

For example a constant fitness function:

def goal(v,k,**kwargs):
    """v is input, either voltage, frequency, time... depends on spice
    simulation. k is the name of node being measured"""
    return 1

Another fitness function: A spice simulation with more than one measurement and example use of **kwargs. This example has a current and voltage measurements:

def goal(v,k,**kwargs):
    if k[0]=='v':   #Voltage
        return kwargs['extra_value'][0]    #1V
    elif k[0]=='i': #Current
        return 1e-3 if kwargs['generation]<10 else 0

Discontinuous fitness function that is 0 before time is 100ns and 5 after it:

def goal(t,k,**kwargs):
    #t is now time
    return 0 if t<=100e-9 else 5

Fitness weights are either constants that scores are multiplied with, or functions that individual values of spice simulation are weighted with, if you want to for example weight some part of the results higher than other parts. They should be in list of dictionaries where number of elements in list is same as number of spice simulations. The nth element is used with nth spice simulation. Keys of the dictionaries are names of measurements(eg. 'v(n2)') and values are weights.

For example a constants weights with 4 simulations with second one having two measurements:

fitness_weight=[{'v(n2)':1},{'v(n2)':1,'i(vin)':50},{'v(n2)':1},{'v(n2)':0.05}]

See examples folder for examples. Due to the rapid development examples might not always work, but I try my best to keep them up to date.