ROLLO Checkpoint File

The checkpoint file is a serialized pickle file and is based on the DEAP checkpointing system.

The checkpoint file holds the ROLLO simulation results and acts as a restart file.

Restarting Simulation

The checkpoint file updates after each generation runs successfully. If the ROLLO simulation fails during a generation, the user can restart the simulation from the previous generation. Therefore, if you’re running a ROLLO simulation with 10 generations and it fails during generation 3, you can restart the ROLLO simulation from generation 2.

This capability can also be used for running one generation at a time. For example, you’re running a large optimization simulation across many nodes on a supercomputer, and don’t want to waste compute time if the simulation fails part way through. You can run one generation at a time, increasing the number of generations defined in the input file with each run.

Results Analysis

To un-pickle the data and analyze the results:

import pickle
checkpoint_file="checkpoint.pkl"
with open(checkpoint_file, "rb") as cp_file:
  cp = pickle.load(cp_file)

Once loaded into Python, the checkpoint file is a dict. The following table describes the keys in the checkpoint file:

Checkpoint File Key	Description
input_file	a copy of the input file
evaluator_files	a dict with copies of the evaluators’ scripts
population	the final population (list of reactor model individuals)
generation	number of generations (int)
hall of fame	deap.tools.HallOfFame object (see deap documentation)
logbook	deap.tools.Logbook object (see deap documentation)
rndstate	simulation’s random state (used when restarting simulation)
all [ind_naming]	dict, key: input parameter name, value: index, corresponds to each individual’s input parameter index in all [populations]
all [oup_naming]	dict, key: output parameter name, value: index, corresponds to each individual’s output parameter index in all [outputs]
all [populations]	list of populations, each population list contains a list of reactor models individuals
all [outputs]	list of population outputs, each population list contains a list of reactor model individual’s output parameters

Each reactor model individual is an Ind class type with attributes and is a simple list (DEAP Individual class). The following code snippet demonstrates a reactor model individual’s attributes:

In [1]:

ind1 = cp["all"]["populations"][0][0]
print(ind1)

[1]:

[7.930883654471881]

In [2]:

print(ind1.__dict__)

[2]:

{'fitness': deap.creator.obj((-7.930883654471881,)),
'gen': 0,
'num': 0,
'output': (7.930883654471881, 1.4598642651422447)}

Descriptions of the reactor model individual’s attributes:

Attribute	Description
fitness	fitness tuple holds the objective values. The sign refers to whether the objective is to maximize or minimize.
gen	generation
num	reactor model index in generation
output	tuple of reactor model individual output parameters

Examples of how to analyze ROLLO results can be found in the Example Notebooks.