automl_zero/experiment.proto

// Copyright 2022 The Google Research Authors.
//
// Licensed under the Apache License, Version 2.0 (the "License");
// you may not use this file except in compliance with the License.
// You may obtain a copy of the License at
//
//     http://www.apache.org/licenses/LICENSE-2.0
//
// Unless required by applicable law or agreed to in writing, software
// distributed under the License is distributed on an "AS IS" BASIS,
// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
// See the License for the specific language governing permissions and
// limitations under the License.


syntax = "proto2";

package automl_zero;

import "fec_cache.proto";
import "generator.proto";
import "instruction.proto";
import "mutator.proto";
import "task.proto";
import "train_budget.proto";


enum FitnessCombinationMode {
  UNKNOWN_FITNESS_COMBINATION = 0;
  MEAN_FITNESS_COMBINATION = 1;
  MEDIAN_FITNESS_COMBINATION = 3;
}

// Stores the entire configuration of an experiment.
message SearchExperimentSpec {
  //////////////////////////////////////////////////////////////////////////////
  // Search tasks. /////////////////////////////////////////////////////////////
  //////////////////////////////////////////////////////////////////////////////

  // Tasks for T_search.
  optional TaskCollection search_tasks = 20;  // Required.

  //////////////////////////////////////////////////////////////////////////////
  // Search space. /////////////////////////////////////////////////////////////
  //////////////////////////////////////////////////////////////////////////////

  // Allowed ops in each of the component functions. To prevent evolving one
  // component function, leave the corresponding list of ops as empty.
  repeated Op setup_ops = 34;
  repeated Op predict_ops = 35;
  repeated Op learn_ops = 36;

  // Initial number of instructions in the component functions. For example, if
  // initial_population=NO_OP_ALGORITHM and predict_size_init=5, the predict
  // component function of every algorithm in the initial population will have
  // 5 no-op instructions.
  optional int32 setup_size_init = 5;    // Required.
  optional int32 predict_size_init = 6;  // Required.
  optional int32 learn_size_init = 7;    // Required.

  //////////////////////////////////////////////////////////////////////////////
  // Search evaluation. ////////////////////////////////////////////////////////
  //////////////////////////////////////////////////////////////////////////////

  // Training budget. If omitted, no training budget is imposed.
  optional TrainBudgetSpec train_budget = 3;

  // Maximum absolute error. If an algorithm produces errors larger
  // than this after any example during in its training or validation, it "
  // will be assigned the minimum fitness if early stopping is used.
  optional double max_abs_error = 14 [default = 100.0];

  // If not present, cache is disabled.
  optional FECSpec fec = 2;

  optional FitnessCombinationMode fitness_combination_mode = 1
      [default = MEAN_FITNESS_COMBINATION];

  //////////////////////////////////////////////////////////////////////////////
  // Search method. ////////////////////////////////////////////////////////////
  //////////////////////////////////////////////////////////////////////////////

  // The population size that should be used by workers that have this dataset.
  // Should the same worker have contradicting values, the code will crash.
  // For each worker, at least one of the datasets must have this field set.
  optional int64 population_size = 33;

  // The tournament size.
  optional int64 tournament_size = 19;

  // The algorith to use to initialize the population. Typically, a
  // NO_OP_ALGORITHM or a RANDOM_ALGORITHM.
  optional HardcodedAlgorithmID initial_population = 4
      [default = NO_OP_ALGORITHM];

  // Total number of train steps executed in the experiment. The default yields
  // ~1000000 individuals each trained for 8000 steps on 10 tasks, when
  // techniques that reduce train steps per individual, such as FEC and hurdles,
  // are not used. Note, "~1000000 individuals" is approximate because of
  // early stopping; however, in practice, we did not witness early stopping
  // drastically reduce the number of individuals.
  optional int64 max_train_steps = 15 [default = 80000000000];

  // The mutation types that can happen during the experiment.
  optional MutationTypeList allowed_mutation_types = 17;  // Required.

  // Probability that a child will be mutated from the parent. If not mutated,
  // the child is identical to the parent.
  optional double mutate_prob = 18 [default = 0.9];

  // The minimum and maximum allowed number of instructions for the component
  // functions. These are required only if insert/remove mutations are used
  // (otherwise, the number of instructions remains at the setup_size_init,
  // predict_size_init, and learn_size_init values).
  optional int32 mutate_setup_size_min = 8;
  optional int32 mutate_setup_size_max = 9;
  optional int32 mutate_predict_size_min = 10;
  optional int32 mutate_predict_size_max = 11;
  optional int32 mutate_learn_size_min = 12;
  optional int32 mutate_learn_size_max = 13;

  // The period between progress reports to stdout, as the experiment
  // advances.
  optional int64 progress_every = 16 [default = 10000];
}