The rig scheduling problem (RSP) is about allocating well activities across a scarce and expensive fleet of rigs, with the goal of avoiding production-delay losses while making efficient use of resources. It’s been studied in operations research since linear-programming production-planning models first emerged in the early 1960s. In their systematic review, [5] proposed a classification that splits the RSP into four major classes: the drilling rig scheduling problem (DRSP), workover planning with and without routing, field development planning, and resource planning, where rigs are scheduled jointly with auxiliary resources such as offshore support vessels. Our formulation falls under the DRSP class. Specifically, it addresses scheduling drilling activities across a heterogeneous rig fleet under operational constraints, treated as an isolated decision separate from broader field development.

Within these RSP classes, mathematical programming forms an important strand of the literature, alongside a large body of heuristic and metaheuristic methods. Mixed-integer linear programming (MILP) formulations have been applied to workover rig scheduling with heterogeneous fleets, release dates, and rig-eligibility constraints, including arc-time-indexed formulations solved through branch-price-and-cut [6]. Related work on scheduling general maintenance tasks for oil and gas production assets has extended discrete-time MILP formulations to multitask work orders with precedence relations, parallel heterogeneous resources, and crews with differing capabilities and work shifts [7]. Since service and drilling durations are uncertain, stochastic extensions have also been proposed: deterministic integer-programming models for rig fleet sizing and scheduling have been paired with two-stage stochastic programs that optimize fleet size under intervention-time uncertainty [8]. Despite these advances, a persistent theme is that realistic instances scale poorly for exact solvers, motivating a large body of heuristic and metaheuristic methods [6].

Alongside optimization-based methods, data-driven approaches have also emerged, using historical records through text mining and regression to produce realistic estimates of intervention durations as inputs to optimization models [6]. A recognized gap in this literature is the scarcity of methods validated in real-world settings: a systematic review found that only nine of the surveyed studies had actually been implemented in practice, which limits collaboration between academia and industry [5]. Heuristic optimization has also been applied to drilling schedule generation, where [9] proposed two algorithms, a greedy economic-selection procedure and a search-space-reduction method, whose schedules achieved net present values exceeding at least 95% of randomly generated schedules and outperformed conventional economic-indicator well ranking. More recently, machine-learning approaches have started to appear as well, including a Markov-chain model that infers rig capabilities from historical rig movements and computes transition probabilities between well classes to automate drilling schedule generation [10].

Reinforcement learning (RL) extends this shift further by learning sequential decision policies for problems with long decision horizons. RL has been shown to improve net present value (NPV) in well-production control compared to conventional optimization-based strategies, and it has also been applied to geosteering and other sequential drilling decisions under geological uncertainty [11]. Closest to our setting, Cárdenas Pantoja et al. [1] formulate drilling-schedule optimization as a Markov decision problem and solve it using a Deep Q-Network augmented with a soft-update mechanism, reporting higher and more stable NPV than a standard DQN baseline on a synthetic field-development case under subsurface uncertainty. These studies establish that learned policies work well for petroleum scheduling, but they remain entirely classical: none of them employs a quantum or hybrid quantum-classical value function, and, importantly, none systematically reports the across-seed reliability of the resulting policies, which is the specific gap our work targets.

The Deep Q-Network (DQN) showed that a neural network trained with experience replay and a periodically updated target network can approximate the action-value function stably enough to reach human-level control on high-dimensional tasks [2]. DQN is a natural fit for scheduling because the decision at each step is simply assigning a specific rig to a specific well, or waiting. That said, value-based deep RL is well known to be sensitive to random seeds, initialization, and hyperparameters. Independently seeded runs can converge to policies of markedly different quality, and reporting only the single best run substantially overstates what typical deployed performance actually looks like. This reproducibility concern is exactly why rigorous evaluation trains across multiple seeds and reports the full distribution of outcomes (mean, variance, and worst case) rather than just the maximum. Our protocol follows this practice, treating variance as a primary object of investigation rather than a nuisance to be averaged away. A related challenge is the large discrete action space that’s typical of assignment problems: at any given step, most rig-well pairs are infeasible, whether because the rig is busy, in maintenance, in the wrong area, or the well itself is unavailable. Simply penalizing invalid actions floods the agent with negative reward and destabilizes learning. Action masking, which restricts the policy to only the currently feasible actions, is a standard remedy that improves both stability and sample efficiency in large discrete action spaces, and we adopt it as the interface between our environment and both agents, regardless of their underlying value function.

Variational quantum circuits (VQCs), which are parameterized quantum circuits whose gate angles are optimized classically, are the dominant model of quantum machine learning in the noisy intermediate-scale quantum (NISQ) era, where they act as function approximators in place of neural networks. Within value-based RL, Chen et al. were the first to use a VQC as the action-value approximator in a DQN, learning discrete environments such as FrozenLake with far fewer trainable parameters than an equivalent classical network [3]. Lockwood and Si extended hybrid VQC agents to continuous state spaces such as CartPole [4], and later attempted to scale a hybrid neural-network/VQC model to Atari environments, where the limited results exposed just how difficult it is to encode large observation spaces and to attribute learning between the classical and quantum components [12]. Skolik, Jerbi, and Dunjko refined value-based quantum agents further with data re-uploading and trainable output scaling on CartPole and FrozenLake, showing that the structured, low-dimensional parameter space of the circuit plays a central role in its behavior, and that output-range scaling is essential for value-based VQCs [13]. On the policy-gradient side, Jerbi et al. introduced parametrized quantum policies and analyzed their expressivity, and in certain special cases established provable separations from classical policies [14]. Work in this area continues on circuit design and optimization for QRL, including QCNN-inspired, noise-robust circuit variants for quantum deep Q-learning that improve robustness under device noise while adopting repeated-run evaluation to stabilize comparisons across random seeds [15].

A central theoretical question for these models is the trade-off between expressibility and trainability. Highly expressive, deeply entangled circuits tend to fall into barren plateaus, regions where gradients vanish exponentially as system size grows, which impedes trainability [16]. This isn’t just an empirical observation either; Holmes et al. formally connected the two properties, showing that the more expressive an ansatz is, the flatter its cost landscape becomes, and therefore the harder it is to train [17]. On the other hand, the restricted and structured hypothesis class defined by a shallow variational quantum circuit can actually support favorable generalization. Caro et al. proved that quantum models can generalize from relatively few training samples, with error bounds that scale with the number of trainable gates rather than the dimension of the Hilbert space [18]. This tension, where too much capacity harms trainability and too little harms expressivity, yet a compact circuit can still generalize well, is precisely the lens through which we interpret our results: a small, 12-parameter circuit turns out to be expressive enough to represent good scheduling value functions while remaining constrained enough to yield policies that stay consistent across random initializations.

Across these threads, three observations stand out. First, classical rig scheduling is dominated by mathematical programming and, more recently, classical RL, none of which examines quantum value functions or policy reliability. Second, DQN’s seed sensitivity is widely acknowledged, but it’s usually mitigated procedurally rather than treated as a target for improvement in its own right. Third, VQC-based agents are almost always evaluated on standard control benchmarks, and their prospective advantage tends to be framed in terms of parameter or sample efficiency rather than robustness. Our work departs from this on all three counts: we apply a hybrid quantum DQN to a realistic, constraint-heavy drilling-scheduling problem, we adopt a controlled multi-seed protocol that makes across-seed variance a primary metric, and we identify variance reduction, meaning reliable completion of the field regardless of initialization, as the main practical benefit of the quantum value function, grounding this claim in the generalization and expressibility theory discussed above. To our knowledge, this is the first application of hybrid quantum reinforcement learning to drilling-rig scheduling.

The field contains 24 wells, partitioned into three geographic areas (A, B, C) of eight wells each. A fleet of six rigs (two per area) drills the wells over a planning horizon of T = 80 days. Each well has a location, a drilling duration of 3 to 6 days, a release day (its earliest permissible start), an integer priority from 1 to 4 (higher is more valuable), a soft due day, and a hard deadline equal to the horizon T. Each rig has a travel speed, a daily operating cost, an assigned area, and a set of scheduled maintenance (down) days. A depth attribute is retained per well for descriptive purposes only and is not enforced as a constraint.

The observation on day t exposes every quantity on which feasibility and future cost depend. It comprises the normalized planning day; five values per rig (an availability flag equal to 1 if the rig is free and not in maintenance, the normalized remaining busy time, and the normalized x and y position of the rig, together with a maintenance flag equal to 1 on a scheduled down day); one occupancy flag per operating area, equal to 1 while the shared crew in that area is committed; and one status value per well. Well status is encoded on a single channel taking the value 0 for a well not yet started, 0.5 for a well whose drilling is in progress, and 1 for a well whose drilling has been completed, so that a committed well still being drilled is distinguishable from one not yet begun. The state dimension is:

This gives the agent a fully observable summary of field progress: rig locations and remaining busy times, on which feasibility and future mobilization costs depend, are included directly in the observation.

The action space contains one assignment action for every rig-well pair, plus one idle action:

Because most rig-well pairs are infeasible on any given day, the environment exposes a Boolean feasibility mask of length 145 indicating which actions are currently valid, and the agent selects only among feasible actions. This action-masking interface, standard for large discrete action spaces in which many actions are invalid [2], prevents the agent from being flooded with invalid-action penalties and is applied identically to both agents compared in this research.

Eight operational constraints govern feasibility and dynamics:

1) Area: A rig may drill only wells located in its assigned area.

2) Release: A well cannot start before its release day.

3) Maintenance: Each rig has scheduled down days on which it is out of service. An assignment is infeasible if the rig is in maintenance on the decision day, or if any down day falls within the drilling interval; the start is otherwise deferred to the earliest clear window. Travel and rig-up days are not tested against maintenance.

4) Weather: No job may start on a weather-blocked day, the start is deferred to the next clear day. Once started, a job may span such days.

5) Shared crew: At most one active drilling job is permitted per area at any time, reflecting a shared crew or key equipment unit limit within each area.

6) Rig-up time: A fixed setup time of 1 day precedes drilling at a newly assigned well.

7) Costs: Operating cost, travel/delay cost, per-kilometer mobilization plus a fixed move-setup cost, and a daily idle spread cost charged for each free rig.

8) Deadlines: A soft penalty applies to completion after the due day. A well that cannot finish by the hard deadline T is infeasible to assign and exactly by the action mask.

When a feasible assignment of a rig to a well is selected on day t, the environment first computes the travel time from the rig’s current location to the well:

The earliest start day accounts for travel, rig-up time, the rig’s current availability, and weather:

where the resulting interval would overlap a scheduled down day of the assigned rig, the start is deferred to the earliest day admitting a clear window of length duration, so a drilling job is never interrupted by maintenance. The well completion day is:

Here, next_clear_day() advances the start past any weather-blocked day, and busy_until is the day the rig becomes free. After assignment, the rig’s busy_until marker is set to end, its location is updated to the well, and the well is recorded as committed. A well is marked complete only once its drilling interval has elapsed, that is, on the first planning day d ≥ end. Assignment and completion are therefore distinct events: between the assignment day and end the well is in progress, is excluded from the feasible action set, and does not yet count toward the completion total. Assignments are irrevocable, and the simulated day then advances by one. Constraints are checked over intervals rather than at the decision day alone. The busy interval recorded for a rig runs from the assignment day to completion and spans travel, rig-up, and drilling, so the shared-crew constraint permits at most one commitment per area throughout. Maintenance is evaluated over the drilling interval, and weather applies only to the start day. Because the same routine computes the interval for both the action mask and the state transition, the mask cannot admit an action that the transition would then execute in violation of a constraint; a separate audit routine re-verifies every realized schedule and reports no violations for any schedule in this paper. An episode terminates once every well has completed or the horizon T is reached. Wells still in progress at the horizon are not counted as complete, and the hard-deadline check guarantees that all committed jobs finish within T.

Each feasible assignment yields the immediate reward:

With the parameter values:

In addition, each free rig incurs a daily idle spread cost:

At episode termination, a proportional completion bonus is awarded:

The proportional form rewards partial progress rather than only a perfect schedule, and is intended to keep the marginal value of drilling each well positive so that completing the field is encouraged. We do not, however, prove that every feasible well yields a positive marginal return under all schedules and parameter settings, so full-field completion is not a guaranteed property of the reward-optimal policy. Feasibility of full completion is established by the heuristic planner in Section 3.7 rather than implied by the reward structure. In practice, this shaping largely reduces the problem to sequencing: the agent decides which feasible well to drill next, and when, to minimize operating, mobilization, idle, and lateness costs. Scheduling quality is therefore measured jointly by the number of wells completed and the total reward.

The agent seeks a policy that maximizes the expected episodic return, defined as the sum of per-step rewards over an episode:

To ensure the problem is solvable rather than merely difficult, a constraint-aware heuristic planner establishes feasibility by assigning, in each area, the free rig to the highest-priority available well. For the problem instance studied in this paper, this planner completes all 24 wells with a makespan of 64 days. We set the horizon to:

The instance is thus provably feasible: a sufficiently capable policy can complete every well within the horizon. Full completion of the field is therefore attainable, and the completion total is a well-defined quantity against which scheduling policies can be compared.

Both agents are Deep Q-Networks (DQN). A Q-network estimates the action-value function Q(s, a), which represents the expected return obtained by taking action a in state s. The agent acts greedily with respect to Q among the currently feasible actions. Training uses standard DQN machinery: an experience-replay buffer, a target network updated by soft (Polyak) averaging, and epsilon-greedy exploration restricted to feasible actions.

Using masked action selection, given state s and feasibility mask m, the local Q-network produces Q-values for all 145 actions. Infeasible actions are set to negative infinity so they can never be selected. With probability (1 - epsilon), the agent picks the highest-valued feasible action; with probability epsilon, it picks a random feasible action. This confines both exploration and exploitation to valid rig-well assignments.

Masked temporal-difference target. For a sampled transition (s, a, r, s_next, done) with next-state mask m_next, the bootstrap target is:

Infeasible next-state actions are masked out before the maximum is taken. The local network is updated by minimizing the mean-squared error between its prediction and this target:

where gamma is the discount factor and Q_target is the target network. Rewards are divided by a fixed scale factor before entering the buffer to keep regression targets numerically stable.

Figure 1 illustrates the architectures of the classical DQN and the proposed hybrid quantum-classical DQN, highlighting the replacement of the classical value-function approximator with a variational quantum circuit in the hybrid model.

Figure 1. Classical and hybrid DQN.

The classical value network is a fully connected multilayer perceptron with the layer sizes:

58 -> 64 -> 64 -> 145

That is, an input layer of size 58, two hidden layers of 64 units each with ReLU activations, and an output layer of size 145 (one Q-value per action). This network contains on the order of 17,300 trainable parameters and serves as the baseline against which the quantum agent is measured.

The hybrid network replaces the hidden representation of the classical MLP with a variational quantum circuit (VQC), while keeping thin classical layers at the input and output. It has three stages arranged as a classical-quantum-classical sandwich.

1) Classical pre-layer. A linear layer maps the 58-dimensional state to 4 values, one per qubit. Each value is squashed by the hyperbolic tangent and scaled by pi, producing four rotation angles in the range [−pi, pi]:

2) Quantum layer. The four angles are processed by a 4-qubit VQC consisting of two parts:

a ZFeatureMap that encodes the four input angles into the qubits as data-dependent rotations (fixed, not trained); anda RealAmplitudes ansatz of trainable single-qubit Y-rotations interleaved with entangling CNOT gates, using two repetitions. The circuit is read out through the Pauli-Z expectation value on each qubit, yielding four real outputs in the range [−1, 1]. The number of trainable quantum parameters and the circuit depth are:

trainable quantum parameters = 12, circuit depth = 10

3) Classical post-layer. A linear layer maps the four measured expectation values to the 145 Q-values: 4 ◊ 145

The VQC is implemented in Qiskit [19] and exposed to the classical training loop through a PyTorch interface, so the entire hybrid network, meaning the classical pre-layer, the quantum parameters, and the classical post-layer, is trained end-to-end by backpropagation using the same masked Q-learning loss as the classical agent. Gradients with respect to the quantum parameters are obtained using the parameter-shift rule via the framework’s estimator. The quantum register runs on a state-vector simulator, with no hardware noise model applied. The idea behind this design is that the quantum core, with only 12 trainable parameters compared to roughly 17,300 in the classical MLP, defines a compact and structured hypothesis class. As discussed in Section 2.3, this kind of constrained class is associated with favorable generalization [18] while still staying shallow enough to avoid the barren-plateau training pathologies that come with highly expressive circuits [16].

To get a fair comparison and to capture reliability rather than just peak performance, both agents are trained and evaluated under an identical protocol.

Using multi-seed training, each agent is trained independently with three random seeds (0, 1, 2), for 30 episodes per seed. All stochastic components, including network initialization, exploration, and replay sampling, are governed by the seed, so the classical and quantum runs differ only in the value network itself.

Once training is complete, each learned policy is evaluated greedily with exploration disabled (epsilon = 0), which ensures deterministic behavior. Starting from the initial field state, the agent selects the highest-valued feasible action at every step until the episode ends. We record two metrics from this greedy rollout: the number of wells completed (out of 24) and the total reward. Since the greedy rollout reflects how the learned policy would actually be deployed in practice, it gives a more faithful measure of policy quality than simply taking the best training episode.

For each agent, we report the mean and standard deviation of greedy-evaluation reward and wells completed across the three seeds, along with the worst-case (minimum) reward. Reporting the full distribution rather than just the single best run matters here, given the known seed sensitivity of value-based deep RL (Section 2.2), and it allows across-seed variance to serve as a primary measure of policy reliability.

All experiments use a realistically synthetic field instance generated with a fixed random seed so that the classical and quantum agents face exactly the same problem. Table 1 shows a summary of the key instance parameters. The instance comprises:

number of wells = 24 (8 per area) number of areas = 3 (A, B, C) number of rigs = 6 (2 per area) planning horizon = 80 days

Well attributes are drawn from fixed ranges (Table 2 shows the full per-well specifications): drilling duration of 3 to 6 days, release day of 0 to 8, priority of 1 to 4, and a soft due day set to the release day plus the duration plus a random offset of 20 to 40 days. The due days therefore span roughly days 26 to 50, which is loose enough that all wells can be completed on time given the shared-crew queue. Each area is served by two rigs with slightly different travel speeds and daily operating costs, and each rig has one or two scheduled maintenance days; Table 3presents the rig speeds, operating costs, and maintenance days. Also, four weather-blocked days are imposed across the horizon (days 9, 10, 23, 24), on which no drilling job may start.

Feasibility of the instance is established by the constraint-aware heuristic planner. The field layout consists of three well-separated clusters, one per area, each with its own pair of rigs positioned at an area base.

Table 1.Drilling-scheduling problem instance summary.

Table 2.Well specifications.

Table 3.Rig specifications.

The environment is implemented as a custom Gym-style simulator in Python. The classical Q-network and the training loop use PyTorch. The quantum layer is built with Qiskit [19] and Qiskit Machine Learning, using a ZFeatureMap for data encoding and a RealAmplitudes ansatz, wrapped so that it can be trained inside PyTorch through the framework’s neural-network connector. Quantum expectation values and their gradients are computed on a state-vector estimator primitive using the parameter-shift rule; no device-noise model is applied, so results reflect the ideal (noiseless) behavior of the circuit.

Both agents are trained under the identical protocol of Section IV-D:

For each seed, the agent is trained for 30 episodes, the best-reward network weights are retained, and the resulting policy is then evaluated greedily with exploration disabled. The quantum agent is executed on the same machine as the classical agent, on the state-vector simulator.

Two primary metrics are reported, both obtained from the greedy (deployment) rollout:

1) Wells completed: the number of wells drilled by the greedy policy, out of 24. This measures whether the policy schedules the full field.

2) Total reward: the sum of per-step rewards over the greedy rollout, including all cost terms and the completion bonus. This measures schedule quality (priority captured minus operating, mobilization, idle, and lateness costs).

For each metric, we report the mean and standard deviation across the three seeds, and for reward, we additionally report the worst-case (minimum) value across seeds. The standard deviation across seeds is treated as the primary indicator of policy reliability.

The constraint-aware heuristic planner completes all 24 wells within a makespan of 64 days, so the horizon is set to 80 days accordingly. The Gantt chart in Figure 5 shows all 24 drilling jobs placed within the horizon while respecting weather-blocked days and rig maintenance, which confirms the instance is solvable by construction. One notable structural feature is that the heuristic schedule only uses one rig per area at a time. This follows directly from the shared-crew constraint, which serializes drilling within each area and leaves the second rig as a backup for maintenance periods. As a result, the agent’s main optimization lever is the timing and ordering of wells within each area, rather than parallel rig operation. The action mask guarantees that every executed assignment satisfies all eight operational constraints, so no policy can produce an infeasible schedule. Differences between methods therefore arise not from constraint satisfaction but from the ordering and timing of assignments, which determine mobilization distance, idle rig-days, and lateness against soft due dates.

Figure 5.Gantt chart of the constraint-aware heuristic planner schedule. All 24 wells are completed within a 64-day makespan, within the 80-day planning horizon.

Both agents were trained for three seeds of 30 episodes each. The overlaid comparison in Figure 4 shows the two training-reward curves interleaving closely throughout, beginning near 9.9 million in the opening episodes, when exploration is close to random, and declining over the training budget to roughly 9.0 million for the classical agent and 9.4 million for the quantum agent, with the quantum curve showing a slightly tighter band in later episodes. Neither curve indicates convergence within the 30-episode budget. Judged by training performance alone, the two agents look comparable, the differences between them appear under greedy evaluation.

Table 4 summarizes the greedy (deployment) evaluation, in which exploration is disabled, and each trained policy is rolled out deterministically.

Table 4.Greedy-evaluation results (3 seeds).

The per-seed results reveal the source of this gap (Figure 3, per-seed bars). The classical agent completes 23 wells on seed 0 but falls to 17 on seed 1 and 11 on seed 2, with reward declining correspondingly from roughly 9.5 million to 6.2 million. The quantum agent completes 22, 22, and 23 wells across the three seeds, with rewards clustered near 9.1 million. Neither agent completes the full field on any seed.

The head-to-head differences (Quantum minus Classical) are:

reward std = classical 1,375,812 vs quantum 202,359

The most significant result is the difference in variance. The across-seed reward standard deviation is 1.38 million for the classical agent and 0.202 million for the quantum agent—a reduction of approximately 6.8 times. The box plot in Figure 6 shows the classical greedy-evaluation reward spanning from roughly 6.2 million to 9.5 million, whereas the quantum reward is compact around 9.1 million. The robustness plot in Figure 7 highlights the worst-case (minimum) reward for each method: the classical worst case is 6.52 million (seed 2), while the quantum worst case is 8.89 million (seed 1). The quantum policy therefore has a substantially higher performance floor.

Figure 6.Distribution of greedy-evaluation rewards across random seeds for the classical and hybrid quantum-classical agents.

Figure 7.Robustness diagram.

To see how the two policies differ in practice, we replayed the best-seed rollout for each agent and broke down its reward. The decompositions are closely matched in structure: priority value, operating cost, and completion bonus are of similar magnitude for both agents, with only a small lateness penalty separating them. The difference between the methods therefore does not lie in the cost efficiency of an individual schedule, but in how consistently each agent performs across seeds.

For completeness, the trained quantum model uses a 4-qubit circuit with 4 encoding parameters (ZFeatureMap), 12 trainable parameters (RealAmplitudes, two repetitions), and a compiled depth of 10 (Figure 8, circuit and architecture diagrams). The full hybrid network—classical pre-layer, quantum core, and classical post-layer—is therefore dominated in parameter count by its two thin classical layers, while the quantum core that shapes the value function’s structure has only 12 trainable parameters, compared with roughly 17,300 in the classical baseline MLP.

Figure 8.Circuit and architecture diagrams.

Summary of Results

In summary: i) the instance is feasible, with all 24 wells schedulable within 80 days by the constraint-aware heuristic planner; ii) neither agent completes the field under greedy evaluation, with the classical agent completing 17.0 wells on average and the quantum agent 22.3; iii) the quantum agent earns 12.9% higher mean reward (9.12M vs 8.08M) and shows lower across-seed reward variance, with a standard deviation of 0.20M against 1.38M, and a higher worst case (8.89M vs 6.24M). Across these three seeds, the quantum value function’s main difference from the classical baseline is consistency rather than peak performance.

The difference between the two agents on this problem lies in consistency rather than peak capability. On their best runs the classical and quantum policies produce schedules of comparable quality, with closely matched reward composition (Section 6). The two methods diverge in how consistently they reach that quality: across three seeds, the classical policy’s greedy-evaluation reward has a standard deviation of 1.38 million against 0.20 million for the quantum policy, and a worst case of 6.24 million against 8.89 million. Neither agent completes the field on any seed, the classical agent averaging 17.0 of 24 wells and the quantum agent 22.3. For an operational scheduling tool a dependable floor matters, but with three seeds and no hypothesis test this difference is a description of these runs rather than an established property of either method.

This tempers what the present results can claim for quantum machine learning on this class of problem. We do not claim quantum superiority, which noiseless small-scale simulation cannot support. Nor do the results yet establish a reliability advantage in any statistical sense. What they show is that a value network built around a 12-parameter variational circuit reached performance comparable to a classical network with roughly 17,300 parameters on this instance, which is a feasibility result for the formulation rather than a demonstration of advantage.

One candidate explanation for the observed spread is the compact and structured hypothesis class defined by the variational quantum circuit. The quantum core has only 12 trainable parameters against roughly 17,300 in the classical multilayer perceptron. A network with that much capacity can fit a high-reward exploratory trajectory closely, and may then produce a deterministic policy that is sharp but brittle. The circuit constrains what the value network can represent, since its parameters enter only as rotation angles on a small entangled register, and this may act as an implicit regularizer. Such an interpretation would be consistent with theory relating the small, structured parameter space of shallow circuits to favorable generalization, where bounds are governed by the number of trainable gates rather than the dimension of the Hilbert space [18], and the circuit is shallow enough (depth 10, four qubits) to remain trainable, avoiding the barren-plateau regime that afflicts deeply entangled circuits [16].

Despite these promising results, several limitations are worth considering when interpreting the findings and assessing how broadly they apply. There are four main ones: the use of simulation rather than real quantum hardware, the evaluation of a single modest-scale instance, the limited number of seeds and training episodes, and the absence of controlled experiments that would pin down the mechanism behind the observed reliability advantage.

1) All quantum computations are performed on a noiseless state-vector simulator. Real quantum hardware introduces gate noise, decoherence, and measurement error, all of which would degrade the circuit’s behavior. The results reported here should therefore be seen as an idealized upper bound on quantum performance, not a hardware-validated one.

2) Our work uses a single synthetic field instance (24 wells, 6 rigs, 3 areas) and a 4-qubit circuit. Whether the reliability advantage holds up across many instances, larger fields, and larger circuits still remains to be shown.

3) Results are reported over three seeds and 30 episodes per seed, without confidence intervals, hypothesis tests, or a formal variance-comparison test. Three seeds are insufficient to establish either a difference in means or a difference in variance between the two methods, so the differences reported in Section 6 should be read as descriptive of these particular runs rather than as statistically established. Seed counts adequate for inference, together with Welch and Brown-Forsythe tests, are left to future work.

4) Model weights are retained for each seed by the highest ε-greedy training reward rather than by a separate greedy validation rollout. A high exploratory return does not imply a strong deterministic policy, so the retained checkpoint may not be the best-performing one, and this choice contributes to the dispersion of the reported results. Implementing validation-based checkpoint selection properly requires held-out validation instances, since selecting on a greedy rollout of the deployment instance itself would be circular.

Future work should extend the research in four directions: i) evaluation across many field instances and larger problem sizes to test whether the reliability advantage scales; ii) execution on noisy simulators and real quantum hardware to assess robustness to device noise; iii) capacity-matched ablations—for example, shrinking the classical network or enlarging the circuit—to test the regularization hypothesis directly; and iv) exploration of alternative encodings and ansatz designs, and of policy-gradient quantum agents, to determine which architectural choices drive the variance reduction.