Benchmarking of Simod v3.6.0 (Fuzzy + Extraneous)

[iharsuvorau]

CC: @marlondumas @orlenyslp @david-chapela

Benchmarking results for v3.6.0 (resource with differentiated fuzzy resource profiles)

Metric names

Short name	Full name
AED	Absolute Event Distance
CAR	Case Arrival Rate
CED	Circadian Event Distribution
RED	Relative Event Distribution
CTD	Cycle Time Distance
NGD	N-Gram Distance (size 2)
NGD(3)	N-Gram Distance (size 3)

Visualization

Raw CSV files with measurements

• Measurements that are visualised above: measurements.csv
• Discovery results folder: results

Configuration

version: 4
common:
  train_log_path: <train_log_path>
  test_log_path: <test_log_path>
  num_final_evaluations: 10
  evaluation_metrics:
    - 3_gram_distance
    - 2_gram_distance
    - absolute_event_distribution
    - relative_event_distribution
    - circadian_event_distribution
    - arrival_event_distribution
    - cycle_time_distribution
  log_ids:
    case: case_id
    activity: activity
    resource: resource
    start_time: start_time
    end_time: end_time
    enabled_time: enabled_time
preprocessing:
  multitasking: false
  enable_time_concurrency_threshold: 0.5
  concurrency_df: 0.75
  concurrency_l2l: 0.9
  concurrency_l1l: 0.9
control_flow:
  optimization_metric: n_gram_distance
  num_iterations: 30
  num_evaluations_per_iteration: 5
  gateway_probabilities: discovery
  discovery_algorithm: sm3
  epsilon:
    - 0.1
    - 1.0
  eta:
    - 0.2
    - 0.6
  replace_or_joins:
    - true
    - false
  prioritize_parallelism:
    - true
    - false
resource_model:
  optimization_metric: circadian_event_distribution
  num_iterations: 40
  num_evaluations_per_iteration: 5
  resource_profiles:
    discovery_type: differentiated_fuzzy
    granularity:
      - 15
      - 60
    fuzzy_angle: 
      - 0.1
      - 0.9
extraneous_activity_delays:
  optimization_metric: relative_event_distribution
  num_iterations: 1

[iharsuvorau]

David thinks including extraneous activity delays discovery together with fuzzy calendars discovery could degrade the results for CTD and RED metrics. We're running now a similar workload but without extraneous delays to confirm.