https://github.com/maxstrange/primordial-ooze
Dirt simple genetic algorithm library for Python
https://github.com/maxstrange/primordial-ooze
genetic-algorithm machine-learning python
Last synced: about 1 month ago
JSON representation
Dirt simple genetic algorithm library for Python
- Host: GitHub
- URL: https://github.com/maxstrange/primordial-ooze
- Owner: MaxStrange
- License: mit
- Created: 2019-02-21T04:56:53.000Z (over 7 years ago)
- Default Branch: master
- Last Pushed: 2019-04-01T21:10:47.000Z (about 7 years ago)
- Last Synced: 2024-04-23T16:09:41.960Z (about 2 years ago)
- Topics: genetic-algorithm, machine-learning, python
- Language: Python
- Size: 41 KB
- Stars: 1
- Watchers: 3
- Forks: 1
- Open Issues: 0
-
Metadata Files:
- Readme: README.md
- License: LICENSE
Awesome Lists containing this project
README
# primordial-ooze
Dirt simple genetic algorithm library for Python
Docs are below.
## Statistics
The Simulation class stores a list `statistics`, which contains instances of the class `Statistics`.
The `Statistics` class contains the following properties:
- `generationidx`, which is the zero-based index indicating which generation this `Statistics` instance is for.
- `maxval`, which is the maximum fitness score for this generation.
- `minval`, which is the minimum fitness score for this generation.
- `avgval`, which is the average fitness score for this generation.
The Simulation class also stores a list `best_agents`, which contains the best agent at each generation.
## Simulation
A GA simulation. The general workflow for this is:
```python
import primordialooze as po
import pandas
import matplotlib.pyplot
sim = po.Simulation(nagents, shape, fitnessfunction)
bestagent, fitness = sim.run()
print("The agent {} has the best fitness, which is: {}".format(bestagent, fitness))
# Dump and plot
fname = "stats.csv"
sim.dump_history_csv(fname)
df = pandas.read_csv(fname)
df = df.drop(['GenerationIndex'], axis=1)
df.plot()
plt.show()
```
### __init__
```python
def __init__(self, population, shape, fitnessfunc, *, seedfunc=None, selectionfunc=None,
crossoverfunc=None, mutationfunc=None, elitismfunc=None, nworkers=0,
max_agents_per_generation=None, min_agents_per_generation=None)
```
#### Args
The following list contains the arguments that are needed. These do not have default values
since the values for these will change dramatically depending on the problem.
- **population**: The number of agents in the first generation. We will generate this many agents
in the initial generation, each of which is a Numpy Array of shape=`shape`.
They will be mutated according to `mutationfunc`, and evaluated each generation
by `fitnessfunc`.
- **shape**: The shape of each agent in the population. Must be a list-like. The shape of the agents
must be a 1D array of whatever length like `(7,)`.
- **fitnessfunc**: The function to use to evaluate the fitness of each agent in the generation.
Must have signature: `fitnessfunc(agent) -> scalar float`. This function
will be evaluated on every single agent in the gene pool at each generation.
If this function is slow, it probably makes sense to use multiprocessing, unless the
gene pool is quite small. See `nworkers`.
#### Keyword Args
These arguments contain (mostly) sensible defaults, but you should definitely make sure these
defaults work for you. You will almost certainly want to change some of these to fit your problem.
- **seedfunc**: The function to use to create the first generation of agents. The function must have
the signature `seedfunc() -> agent of shape 'shape'`. We call this function
`population` times. When `None`, defaults to uniform random
over the range [-1.0, 1.0) in each dimension.
- **selectionfunc**: The function to use to select the agents that are allowed to breed to create the
next generation. Signature must be `selectionfunc(population, evaluations) -> selected_agents`,
where `population` is an n-dimensional array of shape (nagents, agent_length),
`evaluations` is an array of shape (nagents,); `evaluations[i]` contains
the fitness value for `population[i, :]`; `selected_agents` is an n-dimensional array
of shape (nagents_selected, agent_length), which must contain the selected agents.
`population` and `evaluations` are pre-sorted so that `population[0, :]`, corresponds
to `evalutaion[0]` and has the highest evaluation value. Agents which are not selected
are simply discarded, i.e., they will not appear in the next generation (unless randomly
created again as part of crossover/mutation).
If `None`, defaults to selecting the top ten percent.
- **crossoverfunc**: Crossover function to use. Must have signature `crossoverfunc(agents) -> new_agents`,
where `agents` is an n-dimensional array of shape (nselected_agents, agent_length),
and where `new_agents` must be an n-dimensional array of shape (nagents, agent_length).
This function is applied after the selection function is used to determine which
agents will enter the new generation and this function is used exclusively on those
selected agents. Typically, `new_agents` will constitute the entirety of the new generation,
with one exception being if elitism is used (see below) and another exception being
if the mutation function adds new individuals to the gene pool, rather than just mutating
existing ones.
If `None`, defaults to 2-point crossover used on randomly selected pairs from the
breeding agents until `population` agents (or, if `elitismfunc` is None, `0.9 * population`).
- **mutationfunc**: The function to use to apply mutations to the gene pool. The signature must be
`mutationfunc(agents) -> new_agents`, where `agents` is the value returned from
`crossoverfunc` and `new_agents` must be an n-dimensional array of shape (nagents, agent_length).
This function is applied to the result of `crossoverfunc`.
When `None`, defaults to setting each value in 0.05 of the agents to a random value,
where the random value is drawn from a Gaussian distribution of mean = the value being replaced
and stdev = 0.25.
- **elitismfunc**: A function of signature `elitismfunc(generation_index) -> float in range [0.0, 1.0]`.
This function takes the index of the generation (0 for the first generation, 1 for the second, etc.)
and returns the fraction of top-performers to hold over as-is to the next generation.
The elites are duplicated and then, after the new
generation is created via the selectionfunc -> crossoverfunc -> mutationfunc pipeline, they are
reintroduced into the gene pool. This means that if the above pipeline generates 100 agents
and the elitism is set to take 10, the new generation will be composed of 110 agents. If this
is confusing, see `max_agents_per_generation` and `min_agents_per_generation`.
When `None`, defaults to a function that simply returns 0.1 (or 10%) of the gene pool regardless of the
generation.
- **nworkers**: The number of processes to use to parallelize the fitness function. This will default to 0, which will
mean no parallelism at all. `None` will use the number of cores. Otherwise, should be a positive integer.
- **max_agents_per_generation**: The maximum agents to allow into a generation. If the selection, crossover, mutation,
and elitism functions are not handled properly, it is possible for the number of
agents to change per generation. While this may be desired in some circumstances, it
is often not. If this value is negative, we will allow the generations to grow to arbitrary
size. If it is nonzero, after selection, crossover, mutation, and elitism, we will
take all of the candidates as long as they do not number more than this value. If they do,
we take this many at random.
This value defaults to `None`, which means we use `population` as the max.
- **min_agents_per_generation**: The minimum agents to allow making a new generation. If the selection, crossover, mutation,
and elitism functions are not handled properly, it is possible for the number of
agents to change per generation. While this may be desired in some circumstances, it
is often not. If this value is negative or zero, we will allow the generations
to shrink to zero, after which the simulation will stop. If it is nonzero, after selection,
crossover, mutation, and elitism, we will cycle through the candidate agents in random
order, duplicating them until this value is met. Note that we attempt to spread out the
duplication evenly amongst all candidates.
This value defaults to `None`, which means we use `population` as the min.
### run
```python
def run(self, niterations=100, fitness=None)
```
Runs the constructed simulation.
Either runs until `niterations` have passed, or runs until the best fitness is `fitness` or greater.
Returns the best agent along with its fitness.
#### Keyword Args
- **niterations**: The number of iterations to run the simulation to. Defaults to 100. If `None`,
`fitness` will be used (and must not be None). If both this and `fitness` is
specified, we will stop as soon as one or the other condition is met.
- **fitness**: The fitness level to converge on. As soon as one or more agents have this fitness level
or higher, the simulation will stop. Defaults to `None`. If `None` (the default),
`niterations` will be used (and must not be None). If this and `niterations` is
specified, we will stop as soon as one or the other condition is met.
- **printprogress**: If `True` (the default), we will print a progress indication after each generation.
#### Returns
- The agent with the highest fitness score after the simulation ends.
- The fitness of this agent.