In this folder a Genetic Algorithm is implemented with several add-ons that can be user specific added to the baseline to improve performance. Basline simple GA for continuous control tasks.
- using LSTM layer (1D-CNN)
- adding novelty search to the Algorithm
- adding a crossover method to the mutation (two methods to choose from!)
- to run the Algorithm:
python GARNE.py
Flags:
-env : "The environment you want to train the algorithm. Default is Damper-v0"
-ps : "Population size. Default is 1001"
-pc : "Parent count, number of top performer of the current generation that build the new generation. Default is 20"
-g : "Maximum number of generations. Default is 10"
-ls : "Layer size, number of nodes per layer. Default is 10"
-network : "Network type, feed forward (ff), convolutional neural network (cnn) or long-short-term-memory (lstm). Default is ff"
-std : "Noise that is added to the networks weight as mutation. Default is 0.1"
-novelty :"Adds novelty to the algorithm. Default is: False. If choosen, be aware to adapt K_NEIGHBORS value and the behavior characterization (bc in GA_Addon/base.py) depending on the task/environment!"
-crossover : "Adds crossover to the creation process of the new population. The user has two crossover methods to choose from: \nMethod 1: Slices the seeds of both parents in the middle and combines them. keyword argument: 1 \nMethod 2: For each seed in seed_index of both parents. Pick one for the child with a 50/50 prob. keyword argument 2. Default is 0 - No crossover!"
-worker_count : "Numbers of worker that gather training data. Default is 10"
-save_model : "Saving the best model after training in the current directory"
--save_every : "Saving the best Performer after X generations, default is 10"
Example: python GARNE.py -env CartPole-v0 -g 15
Outputs can be monitored with tensorboard:
tensorboard --logdir=runs
If tensorboard is not opening, try with local host flag:
tensorboard --logdir=runs --host localhost
Perfomance of the three different neural network types tested
