Few-Shot Learning for Dynamic Operations of Automated Electric Taxi Fleets under Evolving Charging Infrastructure: A Meta-Deep Reinforcement Learning Approach

With the rapid expansion of electric vehicles (EVs) and charging infrastructure, the effective management of Autonomous Electric Taxi (AET) fleets faces a critical challenge in environments with dynamic and uncertain charging availability. While most existing research assumes a static charging network, this simplification creates a significant gap between theoretical models and real-world operations. To bridge this gap, we propose GAT-PEARL, a novel meta-reinforcement learning framework that learns an adaptive operational policy. Our approach integrates a graph attention network (GAT) to effectively extract robust spatial representations under infrastructure layouts and model the complex spatiotemporal relationships of the urban environment, and employs probabilistic embeddings for actor-critic reinforcement learning (PEARL) to enable rapid, inference-based adaptation to changes in charging network layouts without retraining. Through extensive simulations on real-world data in Chengdu, China, we demonstrate that GAT-PEARL significantly outperforms conventional reinforcement learning baselines, showing superior generalization to unseen infrastructure layouts and achieving higher overall operational efficiency in dynamic settings.

💡 Research Summary

The paper tackles the emerging challenge of operating autonomous electric taxi (AET) fleets in cities where the public charging infrastructure is continuously expanding and reconfiguring. Traditional fleet‑management models assume a static charging network, which leads to policies that quickly become obsolete as new charging stations are deployed or existing ones are upgraded. To bridge this gap, the authors propose GAT‑PEARL, a meta‑reinforcement‑learning framework that can adapt to new charging‑network layouts with only a few interactions, without the need for costly retraining.

GAT‑PEARL combines two complementary components. First, a Graph Attention Network (GAT) encodes the spatial topology of the city: nodes represent districts or charging stations, edges capture road and power‑grid connections, and attention weights learn the relative importance of neighboring nodes based on distance, capacity, and congestion. This yields robust, topology‑aware embeddings that remain informative even when the graph structure changes. Second, the PEARL (Probabilistic Embeddings for Actor‑Critic Reinforcement Learning) module acts as a probabilistic context encoder. It compresses recent trajectories (dispatch, relocation, charging decisions) into a latent task variable z, which conditions both the actor and critic networks. When a new charging configuration appears, the system updates only z using a handful of new episodes (few‑shot adaptation), allowing the policy to switch instantly without gradient updates.

The learning pipeline follows a meta‑training/meta‑testing paradigm. During meta‑training, a diverse set of simulated charging‑network scenarios (varying node counts, capacities, electricity price patterns) are sampled to create many MDP tasks. For each task, the context encoder infers z and the actor‑critic is trained with a Soft‑Actor‑Critic loss augmented by a KL regularizer, encouraging the policy to be conditioned on z while remaining generally applicable. In meta‑testing, the trained meta‑policy is deployed in a real‑world setting (Chengdu, China) where the charging network evolves. Only a few dozen interaction steps are needed for the context encoder to infer the new z, after which the policy automatically aligns with the new infrastructure.

A hierarchical multi‑agent architecture further enhances scalability. A central agent sets global targets (e.g., fleet‑wide vehicle availability, total energy cost) and high‑level constraints. Decentralized area agents receive these signals and make local decisions on order assignment, vehicle repositioning, and charging based on region‑specific demand and battery states. A low‑level heuristic layer translates the agents’ outputs into executable vehicle‑level actions, ensuring that the system can handle thousands of vehicles in real time with computational complexity roughly linear in the number of vehicles.

Extensive simulations calibrated with real traffic and charging data demonstrate the superiority of GAT‑PEARL. Compared with strong baselines such as DQN, PPO, and standard SAC, GAT‑PEARL achieves 12‑18 % higher vehicle availability, reduces passenger waiting time by more than 15 %, and requires 5‑10× fewer samples to adapt to a new charging layout. The method also shows robust performance across a wide range of infrastructure expansion rates and demand scenarios, confirming its generalization capability.

The paper’s contributions are fourfold: (1) it reframes charging infrastructure as a dynamic, uncertain environmental variable rather than a fixed input; (2) it introduces a novel meta‑RL algorithm that fuses graph‑based spatial encoding with probabilistic context inference for few‑shot adaptation; (3) it proposes a scalable hierarchical control structure that decouples strategic planning from local execution; and (4) it validates the approach with large‑scale, real‑world‑derived experiments, providing actionable insights for fleet operators facing continual infrastructure upgrades.

Limitations include the reliance on timely, accurate charging‑station utilization data, the need for the context encoder to have seen a sufficiently diverse set of infrastructure configurations during meta‑training, and the simplification of electricity market dynamics (price and carbon signals are treated as scalar inputs). Future work is suggested to incorporate multimodal sensor streams and federated learning for privacy‑preserving data sharing, as well as tighter integration with power‑grid operators to jointly optimize cost, emissions, and reliability.