ReasAlign: Adaptive Planner-Scheduler

• ReasAlign is an adaptive planner-scheduler framework that integrates autonomous planning with runtime scheduling to resolve resource allocation and operational conflicts in dynamic settings.
• It employs modular components—such as xApps in O-RAN and reinforcement learning agents—to dynamically adjust resources and mitigate conflicts through online policy updates.
• The framework is applied across domains like network management, HPC, safety-critical systems, and production logistics, achieving measurable improvements in throughput and recovery times.

An adaptive planner-scheduler is a composite computational framework that integrates autonomous or semi-autonomous planning with runtime scheduling. This construct dynamically resolves resource allocation, task mapping, or control policy selection problems in environments characterized by context change, uncertainty, or conflicting objectives. It is foundational in domains such as O-RAN network management, high-performance task-parallel computing, neural network program tuning, real-time safety-critical systems, and production logistics. Adaptive planner-schedulers are often RL-driven, agent-based, or use explicit AI-model inference, enabling online adjustment to maintain or improve overall system performance, guarantee constraints, and mitigate operational conflicts (Cinemre et al., 9 Apr 2025, Abduljabbar et al., 2021, Zhang et al., 2022, Alshaer et al., 24 Sep 2025, Novaes et al., 2019, Ma et al., 2018, Tan et al., 2022).

1. Core Architectural Components and System Integration

The canonical adaptive planner-scheduler architecture comprises both planning and scheduling modules, interfaced via policy selectors or conflict detectors.

• Networked RAN Example (O-RAN): Distinguishes between planners (xApps) that each optimize a subset of network control parameters (e.g., power allocation, RBG allocation) and an adaptive scheduler (A2C agent) in the near-real-time RIC. The scheduler receives context (e.g., average user speed, data-arrival rate, KPI targets), activates subsets of xApps (binary activation vector μ), and updates its control policy based on episodic rewards such as normalized throughput. The scheduling window configuration, online enrichment from non-RT RIC via the A1 interface, and dynamic expansion or contraction of the planner pool (with policy head updates) enable zero-downtime adaptivity and conflict mitigation (Cinemre et al., 9 Apr 2025).
• Task-Parallel and Dataflow Systems: The Adaptive Resource-Moldable Scheduler (ARMS) builds a dynamic, platform-independent cost model (indexed by task type and software topology) and selects resource partitions per task at runtime. This moldability allows both scheduling (thread assignment, resource size) and planning (cost anticipation), integrating planning and scheduling in local and global contexts (Abduljabbar et al., 2021).
• AI-Aided Program Tuning and Safety-Critical Systems: In auto-scheduling for neural network compilation (HARL), a hierarchy of MDPs—spanning operation selection, sketch selection, modification operations, and parameter tuning—is coordinated by a controller that allocates search budgets, prunes unpromising configurations, and updates policies via hierarchical RL (Zhang et al., 2022). In safety-critical TTS, offline-trained AI models produce temporal and spatial priorities, used by reconstruction-based schedulers to ensure constraint satisfaction and safe recovery under system mode changes, faults, or preemption events (Alshaer et al., 24 Sep 2025).

2. Formalization of Adaptive Planning and Scheduling

Adaptive planner-schedulers typically employ explicit mathematical formalizations.

• State and Action Spaces: Schedulers operate over high-dimensional state spaces that may encode context variables ($c$), observed KPIs, graph features, program phases, or allocation matrices. Action spaces are often combinatorial, such as all binary vectors over planner activations (O-RAN), execution partition choices (ARMS), or hardware configurations (big.LITTLE scheduling).
• Reward or Cost Functions: Objective metrics include normalized throughput, makespan, workload balance, energy consumption, or task completion rates. In RL configurations, the reward function encapsulates the performance metric and may incorporate penalty terms for conflicts or safety violations.
• Constraint Modeling: Satisfying precedence (e.g., for scheduling dependent jobs or tasks), enforcing collision-freedom in communication messages, or adhering to capacity constraints is implemented via explicit constraints in the problem formulation (Alshaer et al., 24 Sep 2025, Tan et al., 2022).
• Policy and Value Networks: For actor-critic methods, policy $\pi_{\theta}(a|s)$ governs action selection, with value $V_\phi(s)$ estimating expected rewards. Updates follow established RL gradients (e.g., A2C: $$\nabla_\theta J(\theta) = E_{s,a}[\nabla_\theta \log \pi_\theta(a|s) \cdot A_t]$$).

3. Conflict Detection, Handling, and Adaptivity Mechanisms

A defining feature of adaptive planner-schedulers is online conflict detection and resolution across autonomous planners.

• Context-Dependent Conflict Mitigation: In dynamic, multi-planner environments (notably O-RAN), multiple xApps may produce actions that conflict under certain traffic or mobility regimes but are compatible in others. Adaptive scheduling logic exploits episodic and context observations to select non-conflicting planner subsets, retaining prior control actions when temporarily disabling certain planners to avoid resource mismatches (e.g., assigning power to unallocated RBGs) (Cinemre et al., 9 Apr 2025).
• Recovery and Reconstruction in Safety-Critical Systems: A three-pronged reconstruction-based approach in time-triggered systems employs (1) temporal-recovery reconstructor for post-event rescheduling, (2) failure-recovery reconstructor for resource loss handling via context model updating, and (3) full-schedule reconstructor for schedule reassembly after major context shifts (hardware failure or mode change). Safety checks enforce precedence and collision avoidance, while efficient allocation and rapid recovery mechanisms assure low-latency adaptation (Alshaer et al., 24 Sep 2025).
• Adaptive Multi-Agent Negotiation: In supply chain adaptive scheduling, agent-based local optimizers (material and capacity agents) negotiate incremental schedule changes after disruptions. Each agent solves a tractable local MIP for its assignment region, propagating change proposals throughout the BOM, leading to network-wide—but not globally optimal—adaptation in minutes (Tan et al., 2022).

4. Learning Methods and Online Adaptation Strategies

Adaptive planner-schedulers are predominantly RL-driven, but also leverage bandit algorithms and supervised AI inference for policy improvement and exploration-exploitation trade-offs.

• Advantage Actor-Critic (A2C), Q-Learning, PPO: RL approaches update scheduling and planning policies online, incorporating contextual KPIs, stochastic exploration, and entropy regularization. For example, O-RAN scheduling involves online A2C updates every period, with decaying $\epsilon$-greedy exploration. Astro (big.LITTLE) schedules leverage Q-learning to associate static code phases and dynamic hardware states with optimal configurations, refined during runtime (Cinemre et al., 9 Apr 2025, Novaes et al., 2019).
• Hierarchical and Multi-Agent Learning: HARL orchestrates scheduling across neural network operator, sketch, and parameter levels using sliding-window UCB for macro-level budgeting and actor-critic parameter search at the micro level. Parallel exploration tracks are pruned based on cumulative advantage, reallocating resources to the most promising configurations (Zhang et al., 2022).
• Supervised and GNN-Based Selection: For automated planning, graph neural networks select candidate planners using graph embeddings derived from structured task representations. A secondary predictor conditions on initial planner outcome to enable halfway switching if failure is likely, improving coverage over static or offline portfolios (Ma et al., 2018).

5. Empirical Evaluation and Performance Characteristics

Comprehensive evaluation across domains demonstrates the practical impact of adaptive planner-scheduler approaches.

• O-RAN Adaptive Scheduling: Under high-load and mobility, uncoordinated xApps for power and RBG control suffer up to 16% throughput loss; adaptive scheduler recovery reduces loss substantially and can exceed the best individual xApp by up to 4% as the xApp pool is expanded with baseline policies. Scheduler adaptability supports larger xApp sets, multiple objectives, and rapid onboarding of new planners without retraining (Cinemre et al., 9 Apr 2025).
• Resource-Moldable Scheduling (ARMS): Achieves 1.5–3.5× speedup over work-stealing baselines in low-parallelism or memory-bound regimes by dynamically varying execution partition width by kernel type and DAG locality; converges with baselines at full parallelism (Abduljabbar et al., 2021).
• AI-Aided Real-Time Scheduling: Reconstruction-based adaptivity in safety-critical TTS delivers sub-40 ms (for 50 tasks) full-schedule reconstruction, fault recovery under 16 ms, and maintains deterministic guarantees with sub-linear RAM usage (Alshaer et al., 24 Sep 2025).
• Auto-Scheduling in ML: HARL achieves 6–22% operator speedup and 4.3x search speedup on hard instances compared to flat RL or static schedulers, with up to 9% end-to-end DNN inference time improvements (Zhang et al., 2022).
• Multi-Agent Production Scheduling: Decentralized negotiation after disruption achieves order fulfillment rates of 92–100% and recovers within the disruption timespan, massively outperforming manual war-room replanning (Tan et al., 2022).

6. Scalability, Extensibility, and Limitations

Adaptive planner-schedulers are scalable across several axes: size of planner pools, resource dimensions, objective complexity, and heterogeneity of operational contexts.

• Dynamic Expansion and Modularity: O-RAN and ML scheduling frameworks permit the addition or removal of planners/xApps/operators at runtime by dynamically modifying policy output heads or action spaces. Reconstructed schedulers for TTS can accommodate new task sets and changed communication graphs by reevaluating with up-to-date AI-derived priorities.
• Extensibility to Multiple Objectives: Multi-metric optimization—latency, throughput, energy, fairness—can be simultaneously accommodated by designing corresponding reward/cost terms and safety constraints, as in O-RAN and supply chain settings (Cinemre et al., 9 Apr 2025, Tan et al., 2022).
• Limitations: Assumptions of cost-model stationarity (ARMS), regularity in STA or code-phase partitioning (Astro), and two-stage switch policies (GNN-based planning) may not generalize to highly irregular, non-stationary, or adversarial environments. Online learning may incur convergence delays or suboptimality under rapid context drift; certain domains require stronger formal guarantees of safety or compliance (as in safety-critical systems) (Abduljabbar et al., 2021, Novaes et al., 2019, Ma et al., 2018, Alshaer et al., 24 Sep 2025).

7. Representative Application Domains

• RAN/Telecom Network Slicing: Adaptive, RL-driven xApp activation/scheduling for resource parameter conflict management (Cinemre et al., 9 Apr 2025).
• HPC Task Scheduling: Elastic resource slicing and NUMA-aware threaded mapping using online partition cost models (Abduljabbar et al., 2021).
• ML Operator Tuning: Hierarchical search and real-time configuration of tensor program parameters for inference speedup (Zhang et al., 2022).
• Embedded Real-Time Systems: AI-guided, safety-verified reconstructors for time-triggered, multi-processor task scheduling under dynamic events (Alshaer et al., 24 Sep 2025).
• Heterogeneous SoC Scheduling: Compiler-enhanced RL phase switching and hardware configuration in big.LITTLE SoCs (Novaes et al., 2019).
• AI Planning Portfolio Systems: GNN-based planner selection and online adaptive scheduling for cost-optimal task satisfaction (Ma et al., 2018).
• Production and Logistics: Multi-agent decentralized replanning for large-scale BOMs under supply chain disruption (Tan et al., 2022).