Closed schrum2 closed 2 years ago
Although this still produces an error when uncommented (due to lack of graphviz) the actual file with the information for the neural network is now being automatically saved by the fitness batch files. It also seems like the interactive batch files save the population whenever you type 's' into the console. I am not sure there is much more to save after that?
Now there is a batch file for novelty search that automatically saves the archived shapes.
Let's not consider this closed until the graphiviz error is fixed. Please explain how to recreate the error and post the complete error trace here.
The graphviz error is caused by the line (line 206 for me):
visualize.draw_net(config, stats.best_genome(), view=True, filename='{}/{}{}/champion_neural_network.gv'.format(args.BASE_DIR, args.EXPERIMENT_PREFIX, args.RANDOM_SEED), node_names=node_names)
.
This is the line that saves the champion's neural network. It can be recreated with the following command line parameters:
--INTERACTIVE_EVOLUTION=False --BASE_DIR=friday --EXPERIMENT_PREFIX=dragonfruit --RANDOM_SEED=58 --FITNESS_FUNCTION=type_count --DESIRED_BLOCK=SLIME --BLOCK_LIST_EVOLVES=False --SAVE_FITNESS_LOG=True --EVOLVE_FITNESS=True
The complete error trace:
To avoid the error, a try/except was placed. Since the neural network can be viewed online from the file produced, there is no need to install graphviz. This seems to fix the graphviz problems.
All batch files should automatically save as much data as possible for logging purposes. This means saving the population, the fitness plots, and the final champion network, etc.
Also, talk to @MuellMark about making a new batch file for novelty search