A.I. vs Deterministic algorithms

The French version of Wikipedia defines Artificial Intelligence (AI) as a “set of theories and techniques aimed at creating machines capable of simulating human intelligence“. The English version, on the other hand, describes AI “in its broadest sense, is intelligence exhibited by machines, particularly computer systems. It is a field of research in computer science that develops and studies methods and software that enable machines to perceive their environment and use learning and intelligence to take actions that maximize their chances of achieving defined goals.“

However, I find the French definition particularly lacking, and even the English version falls short of capturing the essence of AI. In my opinion, a more accurate definition would be: “systems that are capable of solving problems without being explicitly given the solution path.” They determine the solution path autonomously based on the data and experience they have been provided. This ability to autonomously discover solutions is what fundamentally distinguishes AI from traditional algorithms.

The term intelligence in AI is less of a concern for me, as long as it doesn’t obscure the core characteristic of these algorithms: just like any other software, their primary purpose isn’t to replicate human intelligence, but to find effective solutions. The value of AI lies in providing solutions that would either necessitate highly complex deterministic algorithms or might be unachievable through deterministic methods altogether.

Take, for example, Waze, which uses the A-Star algorithm to calculate routes. This algorithm is deterministic, designed to find the shortest path from a starting point to a goal within a graph or grid. Although it may mimic certain aspects of human decision-making, it does not qualify as AI.

Deterministic approach / workflow approach

In process management, both algorithmic and AI approaches are available:

• Upon receiving an email,
• Identify the customer using his email address.
• Next, search for the customer data in the CRM.
• Then, retrieve the customer’s history from the ERP.
• Finally, prepare a response.

In this approach, the steps to achieve the goal are clearly outlined and predetermined.

AI Approach / Expert System approach

• Objective: When an email is received, prepare a response.
• Objective: To identify the customer use sender email address
• Constraint: The response must include customer history.
• Constraint: Accessing customer history requires customer data from the CRM.
• Constraint: Accessing CRM requires customer id

In this scenario, the system is given the objectives and constraints, but it autonomously determines the optimal path to achieve the desired outcome.

While the deterministic approach is straightforward and easier to implement, updating or modifying the process can be cumbersome, as it involves redefining steps and reprogramming the workflow. This approach may also be less efficient since it follows a fixed path without optimizing based on the specific context or available data.

In contrast, the AI approach allows the system to autonomously determine the most effective path to achieve the desired outcome, making it adaptable to various situations and inputs. However, implementing an AI system is more complex, requiring advanced technology and expertise to develop and maintain.

The AI system has the potential to scale and handle more complex tasks as it learns and gains more data, leading to potential improvements over time. At a cost.

Building Agents

Traditional agents typically use the workflow approach because it offers a clear, predefined sequence of steps to achieve a specific goal.

In this approach, every task is broken down into distinct stages, where the order of actions and the decision points are explicitly defined by the developers.

There’s a lot of good reasons why this approach is commonly used:

• Simplicity: A workflow-based system is straightforward to design and understand. The entire process is mapped out in advance, making it easy to trace the steps from start to finish. Each step follows a predetermined path, ensuring consistency and repeatability.
• Easier Implementation: For many routine processes, such as customer service workflows or back-office operations, the steps involved rarely change. Implementing a workflow is faster and less complex, as it requires little to no real-time decision-making or adaptation.
• Low Overhead: Workflow-based systems don’t require the complex infrastructure of AI-based systems. Traditional agents can be deployed with relatively simple software architectures that have lower computational and maintenance costs.

However, this rigidity is also a limitation. Any change in process or new data input often requires reprogramming the workflow. As complexity grows, the number of fixed paths increases, making the system less adaptable.

Synapse Postmaster Approach

Synapse Postmaster™ uses the AI approach specifically because because it is designed to process large volumes of data, combining information from both the documents it analyzes and external systems.

The challenge lies in managing dependencies between data. For example,

• To accept a customer request the system needs to verify the customer’s payment history,
• To do this, it must first retrieve customer data from the CRM
• Which requires obtaining the customer ID,
• Getting the customer ID involves understanding the content of the letter
• And, before that, the letter must be read and processed.
• Which had already been done to identify the nature of the request.

Determining whether a specific data is necessary for achieving the goal can be expressed as a simple rule. However, when applied across vast datasets, these simple rules multiply into hundreds, if not thousands, of small workflows or branches.

To manage this combinatorial complexity, Postmaster leverages a model of the data relationships, enabling its inference engine to autonomously find the optimal path through the maze of dependencies. This approach streamlines decision-making and execution, allowing the system to efficiently navigate through the numerous interconnected tasks without predefined workflows.

Lessons Learned from Chess engines history

Traditional chess engines like Stockfish used brute force search algorithms combined with human-designed evaluation functions to analyze millions of possible moves quickly. AlphaZero revolutionized chess engines by leveraging deep learning and neural networks to learn and play chess. Recent versions of Stockfish integrate both brute force and AI.

The future may lie in enhancing traditional deterministic software with AI algorithms to achieve superior performance.