Optimizer using Evolution Strategies


class l2l.optimizers.evolutionstrategies.optimizer.EvolutionStrategiesOptimizer(traj, optimizee_create_individual, optimizee_fitness_weights, parameters, optimizee_bounding_func=None)[source]

Bases: l2l.optimizers.optimizer.Optimizer

Class Implementing the evolution strategies optimizer

as in: Salimans, T., Ho, J., Chen, X. & Sutskever, I. Evolution Strategies as a Scalable Alternative to
Reinforcement Learning. arXiv:1703.03864 [cs, stat] (2017).

In the pseudo code the algorithm does:

For n iterations do:
  • Perturb the current individual by adding a value with 0 mean and noise_std standard deviation

  • If mirrored sampling is enabled, also perturb the current individual by subtracting the same values that were added in the previous step

  • evaluate individuals and get fitness

  • Update the fitness as

    theta_{t+1} <- theta_t + alpha * sum{F_i * e_i} / (n * sigma^2)

    where F_i is the fitness and e_i is the perturbation

  • If fitness shaping is enabled, F_i is replaced with the utility u_i in the previous step, which is calculated as:

    u_i = max(0, log(n/2 + 1) - log(k)) / sum_{k=1}^{n}{max(0, log(n/2 + 1) - log(k))} - 1 / n

    As in the paper: Wierstra, D. et al. Natural Evolution Strategies. Journal of Machine Learning Research 15,

    949–980 (2014).

    where k and i are the indices of the individuals in descending order of fitness F_i

NOTE: This is not the most efficient implementation in terms of communication, since the new parameters are communicated to the individuals rather than the seed as in the paper. NOTE: Doesn’t yet contain fitness shaping and mirrored sampling

  • traj (Trajectory) – Use this trajectory to store the parameters of the specific runs. The parameters should be initialized based on the values in parameters
  • optimizee_create_individual – Function that creates a new individual. All parameters of the Individual-Dict returned should be of numpy.float64 type
  • optimizee_fitness_weights – Fitness weights. The fitness returned by the Optimizee is multiplied by these values (one for each element of the fitness vector)
  • parameters – Instance of namedtuple() CrossEntropyParameters containing the parameters needed by the Optimizer
post_process(traj, fitnesses_results)[source]

See post_process()


See end()


class l2l.optimizers.evolutionstrategies.optimizer.EvolutionStrategiesParameters

Bases: tuple

  • learning_rate – Learning rate
  • noise_std – Standard deviation of the step size (The step has 0 mean)
  • mirrored_sampling_enabled – Should we turn on mirrored sampling i.e. sampling both e and -e
  • fitness_shaping_enabled – Should we turn on fitness shaping i.e. using only top fitness_shaping_ratio to update current individual?
  • pop_size – Number of individuals per simulation.
  • n_iteration – Number of iterations to perform
  • stop_criterion – (Optional) Stop if this fitness is reached.
  • seed – The random seed used for generating new individuals