job-shop-lib 1.1.3

An easy-to-use and modular Python library for the Job Shop Scheduling Problem (JSSP)

JobShopLib

JobShopLib is a Python package for creating, solving, and visualizing job shop scheduling problems (JSSP).

It follows a modular design, allowing users to easily extend the library with new solvers, dispatching rules, visualization functions, etc.

There is a documentation page for versions 1.0.0a3 and onward. See the latest pull requests for the latest changes.

See gnn_scheduler for an example implementation of a graph neural network-based dispatcher trained with PyTorch Geometric.

See this Google Colab notebook for a quick start guide! More advanced examples can be found here.

Installation :package:

JobShopLib is distributed on PyPI. You can install the latest stable version using pip:

pip install job-shop-lib

Key Features :star:

• Data Structures: Easily create, manage, and manipulate job shop instances and solutions with user-friendly data structures. See Getting Started and How Solutions are Represented.

• Benchmark Instances: Load well-known benchmark instances directly from the library without manual downloading. See Load Benchmark Instances.

• Random Instance Generation: Create random instances with customizable sizes and properties. See generation package.

• Multiple Solvers:

  • Constraint Programming Solver: OR-Tools' CP-SAT solver. See Solving the Problem.

  • Dispatching Rule Solvers: Use any of the available dispatching rules or create custom ones. See Dispatching Rules.

• Gantt Charts: Visualize final schedules and how are they created iteratively by dispatching rule solvers or sequences of scheduling decisions with GIFs or videos. See Save Gif.

• Graph Representations:

  • Disjunctive Graphs: Represent and visualize instances as disjunctive graphs. See Disjunctive Graph.
  • Agent-Task Graphs: Encode instances as agent-task graphs (introduced in ScheduleNet paper). See Agent-Task Graph.
  • Build your own custom graphs with the JobShopGraph class.

• Gymnasium Environments: Two environments for solving the problem with graph neural networks (GNNs) or any other method, and reinforcement learning (RL). See SingleJobShopGraphEnv and MultiJobShopGraphEnv.

Publication :scroll:

For an in-depth explanation of the library (v1.0.0), including its design, features, reinforcement learning environments, and some experiments, please refer to my Bachelor's thesis.

You can also cite the library using the following BibTeX entry:

@misc{arino2025jobshoplib,
      title={Solving the Job Shop Scheduling Problem with Graph Neural Networks: A Customizable Reinforcement Learning Environment}, 
      author={Pablo Ariño Fernández},
      year={2025},
      eprint={2506.13781},
      archivePrefix={arXiv},
      primaryClass={cs.LG},
      url={https://arxiv.org/abs/2506.13781}, 
}

Some Examples :rocket:

Create a Job Shop Instance

You can create a JobShopInstance by defining the jobs and operations. An operation is defined by the machine(s) it is processed on and the duration (processing time).

from job_shop_lib import JobShopInstance, Operation


job_1 = [Operation(machines=0, duration=1), Operation(1, 1), Operation(2, 7)]
job_2 = [Operation(1, 5), Operation(2, 1), Operation(0, 1)]
job_3 = [Operation(2, 1), Operation(0, 3), Operation(1, 2)]

jobs = [job_1, job_2, job_3]

instance = JobShopInstance(
    jobs,
    name="Example",
    # Any extra parameters are stored inside the
    # metadata attribute as a dictionary:
    lower_bound=7,
)

Load a Benchmark Instance

You can load a benchmark instance from the library:

from job_shop_lib.benchmarking import load_benchmark_instance

ft06 = load_benchmark_instance("ft06")

The module benchmarking contains functions to load the instances from the file and return them as JobShopInstance objects without having to download them manually.

The contributions to this benchmark dataset are as follows:

• abz5-9: by Adams et al. (1988).

• ft06, ft10, ft20: by Fisher and Thompson (1963).

• la01-40: by Lawrence (1984)

• orb01-10: by Applegate and Cook (1991).

• swv01-20: by Storer et al. (1992).

• yn1-4: by Yamada and Nakano (1992).

• ta01-80: by Taillard (1993).

The metadata from these instances has been updated using data from: https://github.com/thomasWeise/jsspInstancesAndResults

>>> ft06.metadata
{'optimum': 55,
 'upper_bound': 55,
 'lower_bound': 55,
 'reference': "J.F. Muth, G.L. Thompson. 'Industrial scheduling.', Englewood Cliffs, NJ, Prentice-Hall, 1963."}

Generate a Random Instance

You can also generate a random instance with the GeneralInstanceGenerator class.

from job_shop_lib.generation import GeneralInstanceGenerator

generator = GeneralInstanceGenerator(
    duration_range=(5, 10), seed=42, num_jobs=5, num_machines=5
)
random_instance = generator.generate()

This class can also work as an iterator to generate multiple instances:

generator = GeneralInstanceGenerator(iteration_limit=100, seed=42)
instances = []
for instance in generator:
    instances.append(instance)

# Or simply:
instances = list(generator)

Solve an Instance with the OR-Tools' Constraint-Programming SAT Solver

Every solver is a Callable that receives a JobShopInstance and returns a Schedule object.

import matplotlib.pyplot as plt

from job_shop_lib.constraint_programming import ORToolsSolver
from job_shop_lib.visualization import plot_gantt_chart

solver = ORToolsSolver(max_time_in_seconds=10)
ft06_schedule = solver(ft06)

fig, ax = plot_gantt_chart(ft06_schedule)
plt.show()

Solve an Instance with a Dispatching Rule Solver

A dispatching rule is a heuristic guideline used to prioritize and sequence jobs on various machines. Supported dispatching rules are:

class DispatchingRule(str, Enum):
    SHORTEST_PROCESSING_TIME = "shortest_processing_time"
    FIRST_COME_FIRST_SERVED = "first_come_first_served"
    MOST_WORK_REMAINING = "most_work_remaining"
    MOST_OPERATION_REMAINING = "most_operation_remaining"
    RANDOM = "random"

We can visualize the solution with a DispatchingRuleSolver as a gif:

from job_shop_lib.visualization import create_gif, plot_gantt_chart_wrapper
from job_shop_lib.dispatching import DispatchingRuleSolver, DispatchingRule

plt.style.use("ggplot")

mwkr_solver = DispatchingRuleSolver("most_work_remaining")
plot_function = plot_gantt_chart_wrapper(
    title="Solution with Most Work Remaining Rule"
)
create_gif(
    gif_path="ft06_optimized.gif",
    instance=ft06,
    solver=mwkr_solver,
    plot_function=plot_function,
    fps=4,
)

The dashed red line represents the current time step, which is computed as the earliest time when the next operation can start.

[!TIP] You can change the style of the gantt chart with plt.style.use("name-of-the-style"). Personally, I consider the ggplot style to be the cleanest.

Representing Instances as Graphs

One of the main purposes of this library is to provide an easy way to encode instances as graphs. This can be very useful, not only for visualization purposes but also for developing graph neural network-based algorithms.

A graph is represented by the JobShopGraph class, which internally stores a networkx.DiGraph object.

Disjunctive Graph

The disjunctive graph is created by first adding nodes representing each operation in the jobs, along with two special nodes: a source $S$ and a sink $T$. Each operation node is linked to the next operation in its job sequence by conjunctive edges, forming a path from the source to the sink. These edges represent the order in which operations of a single job must be performed.

Additionally, the graph includes disjunctive edges between operations that use the same machine but belong to different jobs. These edges are bidirectional, indicating that either of the connected operations can be performed first. The disjunctive edges thus represent the scheduling choices available: the order in which operations sharing a machine can be processed. Solving the job shop scheduling problem involves choosing a direction for each disjunctive edge such that the overall processing time is minimized.

from job_shop_lib.visualization import plot_disjunctive_graph

fig = plot_disjunctive_graph(
    instance,
    figsize=(6, 4),
    draw_disjunctive_edges="single_edge",
    disjunctive_edges_additional_params={
        "arrowstyle": "<|-|>",
        "connectionstyle": "arc3,rad=0.15",
    },
)
plt.show()

[!TIP] Installing the optional dependency PyGraphViz is recommended.

The JobShopGraph class provides easy access to the nodes, for example, to get all the nodes of a specific type:

from job_shop_lib.graphs import build_disjunctive_graph

disjunctive_graph = build_disjunctive_graph(instance)

 >>> disjunctive_graph.nodes_by_type
 defaultdict(list,
            {<NodeType.OPERATION: 1>: [Node(node_type=OPERATION, value=O(m=0, d=1, j=0, p=0), id=0),
              Node(node_type=OPERATION, value=O(m=1, d=1, j=0, p=1), id=1),
              Node(node_type=OPERATION, value=O(m=2, d=7, j=0, p=2), id=2),
              Node(node_type=OPERATION, value=O(m=1, d=5, j=1, p=0), id=3),
              Node(node_type=OPERATION, value=O(m=2, d=1, j=1, p=1), id=4),
              Node(node_type=OPERATION, value=O(m=0, d=1, j=1, p=2), id=5),
              Node(node_type=OPERATION, value=O(m=2, d=1, j=2, p=0), id=6),
              Node(node_type=OPERATION, value=O(m=0, d=3, j=2, p=1), id=7),
              Node(node_type=OPERATION, value=O(m=1, d=2, j=2, p=2), id=8)],
             <NodeType.SOURCE: 5>: [Node(node_type=SOURCE, value=None, id=9)],
             <NodeType.SINK: 6>: [Node(node_type=SINK, value=None, id=10)]})

Other attributes include:

• nodes: A list of all nodes in the graph.
• nodes_by_machine: A nested list mapping each machine to its associated operation nodes, aiding in machine-specific analysis.
• nodes_by_job: Similar to nodes_by_machine, but maps jobs to their operation nodes, useful for job-specific traversal.

Resource-Task Graph

Introduced in the paper "ScheduleNet: Learn to solve multi-agent scheduling problems with reinforcement learning" by Park et al. (2021), the resource-task graph (orginally named "agent-task graph") is a graph that represents the scheduling problem as a multi-agent reinforcement learning problem.

In contrast to the disjunctive graph, instead of connecting operations that share the same resources directly by disjunctive edges, operation nodes are connected with machine ones.

All machine nodes are connected between them, and all operation nodes from the same job are connected by non-directed edges too.

from job_shop_lib.graphs import (
    build_complete_resource_task_graph,
    build_resource_task_graph_with_jobs,
    build_resource_task_graph,
)
from job_shop_lib.visualization import plot_resource_task_graph

resource_task_graph = build_resource_task_graph(instance)

fig = plot_resource_task_graph(resource_task_graph)
plt.show()

---

The library generalizes this graph by allowing the addition of job nodes and a global one (see build_resource_task_graph_with_jobs and build_resource_task_graph).

Gymnasium Environments

The SingleJobShopGraphEnv allows to learn from a single job shop instance, while the MultiJobShopGraphEnv generates a new instance at each reset. For an in-depth explanation of the environments see chapter 7 of my Bachelor's thesis.

from IPython.display import clear_output

from job_shop_lib.reinforcement_learning import (
    # MakespanReward,
    SingleJobShopGraphEnv,
    ObservationSpaceKey,
    IdleTimeReward,
    ObservationDict,
)
from job_shop_lib.dispatching.feature_observers import (
    FeatureObserverType,
    FeatureType,
)
from job_shop_lib.dispatching import DispatcherObserverConfig


instance = load_benchmark_instance("ft06")
job_shop_graph = build_disjunctive_graph(instance)
feature_observer_configs = [
    DispatcherObserverConfig(
        FeatureObserverType.IS_READY,
        kwargs={"feature_types": [FeatureType.JOBS]},
    )
]

env = SingleJobShopGraphEnv(
    job_shop_graph=job_shop_graph,
    feature_observer_configs=feature_observer_configs,
    reward_function_config=DispatcherObserverConfig(IdleTimeReward),
    render_mode="human",  # Try "save_video"
    render_config={
        "video_config": {"fps": 4}
    }
)


def random_action(observation: ObservationDict) -> tuple[int, int]:
    ready_jobs = []
    for job_id, is_ready in enumerate(
        observation[ObservationSpaceKey.JOBS.value].ravel()
    ):
        if is_ready == 1.0:
            ready_jobs.append(job_id)

    job_id = random.choice(ready_jobs)
    machine_id = -1  # We can use -1 if each operation can only be scheduled
    # on one machine.
    return (job_id, machine_id)


done = False
obs, _ = env.reset()
while not done:
    action = random_action(obs)
    obs, reward, done, *_ = env.step(action)
    if env.render_mode == "human":
        env.render()
        clear_output(wait=True)

if env.render_mode == "save_video" or env.render_mode == "save_gif":
    env.render()

Contributing :handshake:

Any contribution is welcome, whether it's a small bug or documentation fix or a new feature! See the CONTRIBUTING.md file for details on how to contribute to this project.

License :scroll:

This project is licensed under the MIT License - see the LICENSE file for details.

References :books:

• J. Adams, E. Balas, and D. Zawack, "The shifting bottleneck procedure for job shop scheduling," Management Science, vol. 34, no. 3, pp. 391–401, 1988.

• J.F. Muth and G.L. Thompson, Industrial scheduling. Englewood Cliffs, NJ: Prentice-Hall, 1963.

• S. Lawrence, "Resource constrained project scheduling: An experimental investigation of heuristic scheduling techniques (Supplement)," Carnegie-Mellon University, Graduate School of Industrial Administration, Pittsburgh, Pennsylvania, 1984.

• D. Applegate and W. Cook, "A computational study of job-shop scheduling," ORSA Journal on Computer, vol. 3, no. 2, pp. 149–156, 1991.

• R.H. Storer, S.D. Wu, and R. Vaccari, "New search spaces for sequencing problems with applications to job-shop scheduling," Management Science, vol. 38, no. 10, pp. 1495–1509, 1992.

• T. Yamada and R. Nakano, "A genetic algorithm applicable to large-scale job-shop problems," in Proceedings of the Second International Workshop on Parallel Problem Solving from Nature (PPSN'2), Brussels, Belgium, pp. 281–290, 1992.

• E. Taillard, "Benchmarks for basic scheduling problems," European Journal of Operational Research, vol. 64, no. 2, pp. 278–285, 1993.

• Park, Junyoung, Sanjar Bakhtiyar, and Jinkyoo Park. "ScheduleNet: Learn to solve multi-agent scheduling problems with reinforcement learning." arXiv preprint arXiv:2106.03051, 2021.