Inartificial intelligence: is a fool with a large language model login a data scientist?

It Takes a Sophisticated Agent to Use AI Sensibly in Complex Transport Networks

In an era where transport systems are becoming increasingly complex, artificial intelligence (AI) tools such as neural networks and large language models are emerging as pivotal in addressing technical problems. The necessity for highly accurate predictions of what will happen ‘next’ is already a critical component of efficient transport operations. This white paper explores how we can utilise AI most effectively to prepare for future operational models, highlighting the limitations of singular AI models and advocating for an agent-based approach in managing complex systems. 

The Challenges of Applying AI in Complex Transport Systems

AI models have demonstrated remarkable capabilities in handling specific, well-defined tasks within transport networks. For instance, neural networks can predict traffic flow in a particular area, while machine learning algorithms can optimise scheduling for a set of routes or predict the failure of components based on signals. However, these models often perform best when their scope is narrowly focused or trained on identical components. This is not usually the real world environment found in transport networks. Applying a single AI model to manage the entirety of a complex transport network presents significant challenges.

One common misconception is that creating accurate predictions is as simple as ‘training a model’. This oversimplification overlooks the intricacies involved in data preparation, feature selection, model validation, and continuous learning. Technicians turned data scientists may underestimate these complexities, leading to models that underperform or fail when deployed in real-world scenarios.

Moreover, transport networks are dynamic systems with numerous variables and unpredictable events. Singular AI models may struggle to adapt to sudden changes, such as unexpected disruptions, shifts in passenger behaviour, or alterations in data formats. These limitations necessitate a more sophisticated approach to harnessing AI effectively in complex transport environments.

The Power of Agent-Based Approaches

For more intricate challenges, such as managing an entire data pipeline across a large network, a system of AI agents may be more effective than relying on a single model. An agent-based approach involves deploying multiple AI agents, each designed to perform specific tasks or manage particular components of the system. These agents can operate autonomously but also communicate and collaborate to achieve overarching objectives.

This approach enhances flexibility and resilience. If one part of the pipeline encounters an issue or irregularity, the other agents can adjust their operations accordingly, ensuring that the system as a whole continues to function optimally. Additionally, agent-based systems can more readily accommodate changes in data formats or incorporate new data sources without requiring extensive reconfiguration of the entire model.

Strategies for Effective AI Implementation in Transport Networks

To maximise the benefits of AI in transport systems, a combination of advanced modelling techniques and practical considerations is essential.

1. Embracing Agent-Based Models: specialisation helps

Implementing an agent-based system allows for specialised AI models to handle distinct aspects of the transport network. For example, one agent might focus on predicting passenger demand in real time, while another manages route optimisation based on current traffic conditions. This division of labour enables each agent to be fine-tuned for its specific task, improving overall system performance.

2. Leveraging Traditional Models Wisely: linear regression may be your superpower

While advanced AI models receive much attention, traditional methods like linear regression can still play a vital role. Linear regression (LR) often performs well at a high level, providing valuable insights (and very quickly at scale) into trends and relationships within the data. However, practitioners must be cautious of pitfalls such as multicollinearity and over-dimensionality, which can make results appear more favourable than they truly are. Rigorous testing and validation are necessary to ensure the reliability of these models.

3. Avoiding the ‘One-Size-Fits-All’ Mentality: a ChatGPT login does not the data scientist maketh

Recognising that no single model can address all the complexities of a transport network is crucial. Bespoke models or transformations may be needed for different elements of the system and the right model should be chosen for the right use case. Where full disclosure or algorithmic audit is required, LR will be necessary (unless you want different agents to assess the bias in your neural networks).  Customised solutions can cater to the unique characteristics of various data streams, operational requirements, and performance goals. We make this point over single classes of models. In a number of deployment of predictive failure models in railway networks, we have found that identical point machines with identical usage patterns need separately trained machine learning models in order to accurately predict failure. 

4. Continuous Learning and Adaptation

Transport networks are not static; they evolve over time with changes in technology, regulations, and user behaviour. AI systems must be designed with the capacity for continuous learning and adaptation. Regular updates, retraining of models, and incorporation of new data sources help maintain the relevance and effectiveness of AI applications.

Conclusion

As transport systems grow in complexity, the intelligent application of AI becomes increasingly important. While singular AI models excel at specific tasks, managing an entire network demands a more nuanced approach. Agent-based systems offer a promising solution, providing flexibility, resilience, and adaptability. By deploying multiple specialised agents and embracing both advanced and traditional modelling techniques, transport operators can harness the full potential of AI.

However, it is essential to proceed with caution and expertise. Misapplying AI can lead to suboptimal outcomes or even exacerbate existing challenges. Avoiding common mistakes, such as oversimplifying model training or neglecting potential pitfalls in statistical methods, is critical. Ultimately, a sophisticated and strategic approach to AI implementation will position transport networks to meet future operational demands effectively.

How We Have Addressed These Points in Our Platforms

To address these points, we have embraced an agent-based approach within our platforms, deploying multiple AI agents tailored to specific functions such as demand forecasting, route optimisation, and anomaly detection.

This architecture approach allows for greater flexibility and robustness, enabling our platforms to adapt to changes and manage irregularities effectively. Recognising the value of traditional modelling techniques, we incorporate linear regression where appropriate, leveraging the strengths of various analytical methods while vigilantly identifying and mitigating issues like multicollinearity and over-dimensionality to maintain the integrity and reliability of our models.

Understanding that no single solution fits all scenarios, we develop bespoke models and transformations for different elements of the data pipeline, ensuring each component operates optimally within its context and contributes to the overall efficiency of the system. Our platforms are designed for continuous learning, with mechanisms in place for regular model updates and integration of new data sources; collaboration between data scientists, engineers, and transport experts is a cornerstone of our approach, fostering innovation and ensuring our AI solutions are grounded in operational realities. 

By adopting this sophisticated and nuanced application of AI, we are preparing our platforms—and, by extension, our clients—for the near future of transport operations, committed to providing intelligent, adaptable, and effective solutions that harness the full potential of AI while avoiding common pitfalls and oversimplifications.