Managing Oil and Gas Field Development with a Multi-Agent System and Reinforcement Learning

Konstantin Pechko23 May 2026

Problem

An oil company has a backlog of wells to drill and a limited number of crews. The order matters: infrastructure may not be ready, not every crew can reach every well, and a very productive well can break the field's production limit (the "shelf"). Wells interfere, so their effects cannot be ranked separately. The goal is the highest net present value (NPV) over decades. One plan is checked by a field simulation that takes days or weeks, and engineers usually improve a greedy first draft by hand.

Example from the article: a typical field with 300 wells and 20 crews has about 10390 possible plans, so brute force is out of reach.

Solution in brief

Treat planning as a sequence of decisions in a simulated environment. A DQN agent picks the next job and crew, constraints filter infeasible actions, and the digital twin with an economic model scores the plan. The agent starts from weights tuned to copy the greedy plan, so it is no worse than the baseline, and then improves on it, working beside the engineer as a second opinion.

Questions this case answers
  • How to order a backlog of expensive jobs under crew and capacity limits when each candidate plan is evaluated by a slow simulator
  • How to add a learned planner next to expert planning without risking a plan worse than the baseline
  • How to plan a well-drilling backlog when crews, infrastructure and production limits constrain the order of work
  • How to optimize a schedule when each scenario is evaluated by a slow simulator

Tools and methods

  • DQN (deep Q-network) — a neural network scores each available action
  • Hot start from the greedy policy — initial weights reproduce the greedy schedule
  • Digital twin plus economic model — forecast production and convert it to NPV
  • Gymnasium — the interface that wraps the environment
  • Greedy, simulated annealing, brute force — baselines in the experiment
Continue reading — it's free

Google or a one-time email code. No password.