emi-group / evox Goto Github PK

We should move to Gymnasium as Gym is no longer maintained.
Brax is also a good choice.

• Gymnasium
• Brax

A good looking readme is needed.

We need metrics, mostly for multi-objective problems.

• GD
• IGD
• HV

• CMA-ES
• DE differential evolution
• Open-es (openai-es)
• xNES (natural evolution strategy)

Hi,

Thank you for providing this fantastic library. I'm interested in adding a new algorithm, specifically the 'cross-entropy method.'
Could you please provide an example or guidance on how to do this?

Thank you.

Currently, we are using pyproject.toml, update it and add yourselves to the list of authors.

When I initialize the dzlt1
dtlz1 = problems.numerical.DTLZ1()

if parameter m is not given, it would show 'TypeError: unsupported operand type(s) for *: 'int' and 'NoneType''

I think the reason is when you call parent's init funtion, the m was given instead of self.m

• Core
  • State
  • Module
  • Algorithm
  • Problem
• Problems
• Algorithms
• Operators

Found this project from cs.NE; thanks for sharing! 🦎

I cloned the repo; it's too large due to all the blobs committed in git.

Currently 300mb.

For reference:

• JAX: ~67mb
  • EvoJAX: 56mb
  • evosax: 7mb
• NumPy: 117mb

blobs.txt

(For reference, the command to find these):

git rev-list --objects --all \
| git cat-file --batch-check='%(objecttype) %(objectname) %(objectsize) %(rest)' \
| sed -n 's/^blob //p' \
| sort --numeric-sort --key=2 \
| cut -c 1-12,41- \
| numfmt --field=2 --to=iec-i --suffix=B --padding=7 --round=nearest

It might be worth considering rewriting your git history so users can hack around more tractably.

Glad to see more activity in this space! 💪🏻
Thanks for the cool lib!

• t-SNE
• PCA
• treemap?

• More datasets
• More network architectures
• Openai Gym
• EvoXBench

Desciption

In the CEC2022 single objective benchmark, F10 ~ F12 all use the compose_operat function in OperatFunc class, where the codes below may lead to NaN items in term1, resulting in final NaN in the evaluated fitness.

diff_square_sum = jnp.sum((x - Os_mat) ** 2, axis=1)
term1 = 1 / jnp.sqrt(diff_square_sum)
term2 = jnp.exp(-0.5 * diff_square_sum / (sigma**2 * D))
W = term1 * term2
W_norm = W / jnp.sum(W)

Possible Solution

diff_square_sum = jnp.sum((x - Os_mat) ** 2, axis=1)
term1 = 1 / jnp.sqrt(diff_square_sum)
term2 = jnp.exp(-0.5 * diff_square_sum / (sigma**2 * D))
W = term1 * term2
term1_nan = jnp.isnan(term1)
W_norm = jnp.select(jnp.any(term1_nan), term1_nan / jnp.count_nonzero(term1_nan), W / jnp.sum(W))

Location

- problems
  - numerical
    - cec2022_so.py
      - line 121

Is this the intended behavior of CEC2022 single objective benchmark or a bug?

Note

The overall code of CEC2022 benchmark is hard to read due to the lack of documentations and improper abbreviations, please fix these problems if possible.

• Sampling
  • Uniform
  • Latin Hypercube
• Selection
  • Uniform
  • Tournament
• Mutation
  • Polynomial
  • Discrete
• Crossover
  • Uniform
  • Simulated Binary
  • DE (binary/best)
  • One point

Greetings everyone,

I'm currently facing an issue while attempting to run the quick start code. This is leading to an error message: "AttributeError: 'tuple' object has no attribute 'evaluate'".

from evox import algorithms, problems, pipelines
import jax
import jax.numpy as jnp

algorithm = algorithms.PSO(
    lb=jnp.full(shape=(2,), fill_value=-32),
    ub=jnp.full(shape=(2,), fill_value=32),
    pop_size=100,
)

problem = problems.classic.Ackley(),


pipeline = pipelines.StdPipeline(
    algorithm,
    problem,
)

key = jax.random.PRNGKey(42)
state = pipeline.init(key)


for i in range(100):
    state = pipeline.step(state)

Traceback:

---------------------------------------------------------------------------
AttributeError                            Traceback (most recent call last)
/home/sh/test_evox.py in line 27
     25 # run the pipeline for 100 steps
     26 for i in range(100):
---> 27     state = pipeline.step(state)

File ~/anaconda3/envs/EVOX_test/lib/python3.10/site-packages/evox/core/module.py:31, in use_state..wrapper(self, state, *args, **kwargs)
     28 # find the state that match the current module
     29 path, matched_state = state.find_path_to(self._node_id, self._module_name)
---> 31 return_value = func(self, matched_state, *args, **kwargs)
     33 # single return value, the value must be a State
     34 if not isinstance(return_value, tuple):

File ~/anaconda3/envs/EVOX_test/lib/python3.10/site-packages/evox/pipelines/standard.py:24, in StdPipeline.step(self, state)
     21 if self.pop_transform is not None:
     22     pop = self.pop_transform(pop)
---> 24 fitness, state = self.problem.evaluate(state, pop)
     26 if self.fitness_transform is not None:
     27     fitness = self.fitness_transform(fitness)

AttributeError: 'tuple' object has no attribute 'evaluate'

I'm running this code in a clean anaconda env with cuda 12.2.

Best
Sebastian

Assign labels like noisy/multi-objective/single-objective to problems

• NSGA II
• MOEA/D
• RVEA
• IBEA

Hi there! Nice to see more enthusiasm for JAX and evolutionary optimization. Any reason why you don't mention EvoJAX and evosax in your paper? Would love to know how they compare.

The CSO algorithm's generation of new population ends here, and it seems that there is no code like jnp.clip or jnp.min, jnp.max. Is it okay to let the population run free on an unconstrained sapce?

• Parallel Coordinate
• Animation
• 3D plot

Desciption：

When calling the CPSOS algorithm from evoX, the following issues may arise：

ValueError: Incompatible shapes for broadcasting: shapes=[(2, 100), (2,)]

It seems that there is an issue with the cpso_s.py at line 100.
In the comments of this algorithm, the shape of 'fitness' is 'dim, pop_size', which is different from how it's done in other algorithms?

Additionally, there might be a small error in the EvoX tutorial.
This algorithm should be named 'CPSO_S' instead of 'CLPSO_S'

 Algorithms 
    - Single-objective 
       - 【CLPSO_S】

• ZDT
• Rosenbrock
• Griewank
• Rastrigin
• DTLZ

The current non-dominate sort is quite naive and not jit-compatible, we need a better implementation.