EvoRoute: Adaptive Multi-Objective Routing

• EvoRoute is a family of multi-objective routing frameworks that leverage stochastic feedback, domain knowledge, and historical data to optimize trade-offs between travel time, energy consumption, and operational cost.
• In eco-routing applications, EvoRoute’s sub-population assignment algorithm achieves 13-14% energy savings for BEVs and up to 10% travel time reductions under high congestion.
• For EV field operations and agentic LLM systems, EvoRoute integrates hybrid clustering with genetic algorithms and experience-driven policies to reduce route distances by up to 35% and costs by up to 80%.

EvoRoute refers to a family of algorithms and frameworks focused on adaptive, multi-objective routing in complex network environments. The term encompasses methodologies for transport (eco-routing for BEVs/ICEVs), large-scale combinatorial optimization (EV field service logistics), and experience-driven stepwise routing in agentic AI systems. In all cases, EvoRoute’s defining trait is multi-objective trade-off optimization—explicitly leveraging system feedback, domain knowledge, and/or historical records to guide near-optimal assignment, route selection, or subtask scheduling under competing operational criteria.

1. Multi-Objective Eco-Routing for BEVs and ICEVs

In the context of transportation networks, EvoRoute denotes a stochastic, feedback-driven sub-population assignment algorithm aiming to minimize both travel time and energy/fuel consumption for heterogeneous vehicle fleets. The underlying mathematical model considers a directed graph $G=(N,A)$, OD-pair-specific vehicular demands $Q_{rs}^v$ by class (BEV, ICEV), and distinct cost components per link: travel time $T_i(x_i)$ (BPR-type function) and vehicle-specific energy/fuel $E_i^v(x_i)$ computed via VT-Micro (ICEV) and VT-CPEM (BEV) models. Each class $v$ trades off energy versus time with a coefficient $\alpha^v$, yielding a multi-objective user equilibrium (MO-UE) traffic assignment:

$$\min_{f} Z(f) = \sum_{i\in A} \sum_{v=1}^{V} f_i^v \left[ \alpha^v E_i^v(x_i) + (1-\alpha^v) T_i(x_i) \right]$$

subject to flow conservation and nonnegativity. The equilibrium solution corresponds to a class-specific Nash point considering both objectives simultaneously.

The EvoRoute implementation partitions each user class into five equal sub-populations, updating only one per iteration (20% of traffic per demographic class), solving the corresponding shortest-path problems on dynamically estimated link costs augmented with stochastic noise, and updating flows via successive averages ($w_k=1/k$). Oscillation is controlled by this decomposition, while exploration is enhanced through injected noise in link costs. The method converges to a multi-class MO-UE with empirically verified performance in microscopic traffic simulations of the Los Angeles network, supporting over 80,000 OD-pairs with demand up to 140,000 vehicles per hour.

2. Hybrid Clustering–Genetic Algorithm for EV Field Operations

EvoRoute is adapted for the operational planning of EV charging network maintenance, particularly for large cities with complex infrastructure. The principal challenge is minimizing total driving distance/time for technicians tasked with maintaining geographically dispersed stations, a combinatorial instance of the vehicle routing problem (VRP).

The approach first clusters $n$ stations into $k$ spatially compact groups via K-means, minimizing

$$J(\{\mu_i\}) = \sum_{\ell=1}^k \sum_{x_j \in S_\ell} \| x_j - \mu_\ell \|^2$$

followed by solving two levels of TSP: (1) a global tour on cluster centroids using a genetic algorithm (GA), and (2) intra-cluster tours. Chromosomes encode centroid or station permutations; fitness is inverse of route length. Genetic operations include tournament or roulette-wheel selection, order crossover (OX), and swap mutation, with parameters such as mutation rate $p_{mut} \in \{0.05, 0.10\}$ and generations $\in \{100, 200\}$. Constraints enforce connectivity (enter/exit once), tour closure, and subtour elimination (Miller–Tucker–Zemlin).

Empirical evaluation on the Istanbul 100-station dataset shows distance reductions of up to 35% compared with sequential routing, with best configurations ($k=5$, $p_{mut}=0.10$, $100$ iterations) reducing route distance from 417.5 km to 308.5 km. The pipeline is scalable ($k=O(\sqrt{n})$) and modular, supporting extension to larger instances and dynamic factors (e.g., live traffic, time windows), though the baseline implementation uses static Euclidean clustering and fixed routes (Yenigun et al., 22 Sep 2025).

3. Experience-Driven Self-Routing in Agentic LLM Systems

In composite LLM-based agent architectures, EvoRoute is a dynamic, Pareto-optimal model selection and routing policy designed to address what the literature terms the “Agent System Trilemma”: simultaneous maximization of performance $P(\tau)$ (accuracy), minimization of monetary cost $C(\tau)$, and minimization of wall-clock delay $D(\tau)$. The policy $\rho$ must, at each subtask step $t$, choose an LLM $l_t$ from a candidate pool $L$ to maximize expected multi-objective utility:

$$\rho^* = \arg\max_\rho \left( \mathbb{E}_{\tau \sim \rho}[P(\tau)],\, -\mathbb{E}_{\tau \sim \rho}[C(\tau)],\, -\mathbb{E}_{\tau \sim \rho}[D(\tau)] \right)$$

EvoRoute maintains a growing experience base $K$ of prior invocations (including agent, query embedding, tools, cost, delay, and success metrics), and at each decision point:

1. Retrieve historical relevant records based on agent, semantic similarity, and tool usage.
2. Aggregate statistics for available LLMs (mean performance, cost, delay).
3. Pareto-filter to retain only non-dominated LLMs on all three axes.
4. Sample a posterior over each objective (Normal-Inverse-Gamma) and select $l^*$ via Thompson sampling of a utility-weighted scalar.
5. Execute and append the new experience to $K$ for continual policy refinement.

Cold-start exploration uses uniform sampling over the LLM pool, after which all routing is experience-driven. No explicit RL is used; the routing policy evolves via feedback and expanding $K$. The approach generalizes to arbitrary multi-agent workflows by replacing the static agent-to-LLM assignment $\varphi$ with the dynamic pipeline $\rho(s_t, q_t, i_t, K)$ (Zhang et al., 6 Jan 2026).

4. Empirical Evaluation and Key Results

The multi-objective eco-routing EvoRoute implementation demonstrated BEV energy savings of 13.5–14.2% and ICEV fuel savings up to 10.6% across congestion regimes. Notably, under high congestion, average travel times decreased by 10.1% compared to standard user equilibrium, moving closer to system optimum assignments. Tuning $\alpha^v$ calibrates the energy–time trade-off, mapping out the Pareto frontier between objectives. Under single-class assignment, energy savings rise but incur higher time penalties, supporting adaptive policy selection by system designers (Ahn et al., 2020).

For EV field logistics, the Istanbul case found that clustering (e.g., $k=5$) and GA routing provide distance reductions of up to 35% over sequential baselines, with operational duration decreasing from 706 to 568 minutes under best configuration. Clustering effectively reduces combinatorial complexity, with each sub-tour parallelizable (Yenigun et al., 22 Sep 2025).

In LLM agent systems, EvoRoute's experience-driven self-routing reduced execution costs up to 80% and latency over 70% on benchmarks including GAIA and BrowseComp+. For example, on BrowseComp+, EvoRoute+CK-Pro increased accuracy by 5.2 percentage points over a GPT-4 baseline while cutting cost by 64% and duration by 49%. Ablation studies confirm that multi-faceted retrieval, Pareto filtration, and Thompson sampling all contribute significant improvements, with cold-start essential for robust initial performance (Zhang et al., 6 Jan 2026).

5. Algorithmic and Architectural Insights

All EvoRoute instances leverage multi-objective optimization and decomposition. In eco-routing, the key innovation is sub-population feedback assignment with stochastic exploration, updating only a fraction of vehicles at each step to prevent cycling and ensure robust convergence. In field service routing, K-means clustering with GA-based TSP search modularizes the problem, explicitly supporting scalability and parallelization. The agentic LLM system recomputes routing decisions at every subtask using fine-grained, context-sensitive experience selection, with Pareto pruning and Thompson sampling to avoid suboptimal or dominated choices.

Each EvoRoute instantiation explicitly allows for extension: eco-routing supports connected-vehicle V2I/V2V feedback, the logistics variant admits real-time traffic and technician constraints, and the LLM agent system directly integrates into any turn-based workflow with a minimal interface overhead.

6. Limitations and Future Directions

Reported limitations of EvoRoute include the following. In eco-routing, energy–time trade-offs may yield higher average trip times, particularly under low congestion for highly energy-focused settings. The Istanbul field logistics implementation uses static, geographic clustering and does not model road network topology, dynamic traffic, or multi-technician scenarios; richer heuristics and real-time adaptations remain open avenues. In agentic systems, EvoRoute requires sufficient initial exploration and assumes that LLM costs/delays are stable across time and load.

Future research directions include augmenting eco-routing with real-time re-routing and V2I integrations, adapting field service EvoRoutes to time-dependent VRPTW and capacitated, multi-depot scenarios, and incorporating adaptive policy RL or bandit feedback into LLM agent systems for greater long-term adaptivity (Ahn et al., 2020, Yenigun et al., 22 Sep 2025, Zhang et al., 6 Jan 2026).

7. Comparative Summary and Applications

EvoRoute frameworks demonstrate robust, empirically validated improvements across diverse routing settings, as summarized in the comparison below.

Application Domain	Primary Optimization Objectives	Empirical Gains vs Baseline
Eco-routing (urban traffic)	Energy/fuel vs time (multi-class)	Up to 14% BEV & 10% ICEV savings; –10% time
Field logistics (EV maintenance)	Route length, duration	25–35% distance, 20% time reduction
Agentic LLM routing (AI systems)	Task accuracy, cost, latency	Up to 80% cost, 70% latency reduction; +performance

In all implementations, EvoRoute balances competing metrics and leverages system experience or domain structure for efficient, adaptive routing. Its modularity and ease of integration make it suitable for direct application or extension in large-scale, real-world environments across both physical and digital domains (Ahn et al., 2020, Yenigun et al., 22 Sep 2025, Zhang et al., 6 Jan 2026).