AI DSS (Artificial Intelligence Decision Support Systems) (Part 1)

Introduction

A broad concept often encompassing many different fields with various levels of "tolerance", a DSS or Decision Support System is a platform which is as the name suggests designed to perform decisions, or assist with making decisions, there is many cases of DSS's most commonly found in the loans branch of technology. An AI DSS is designed to leverage the capabilities of AI to assist in the process of decision making. An example of this might also include the process for a medication selection lets say two courses of action in a hospital might be provided by a DSS as a way to treat a difficult problem, it provides several candidates for medications based on the inputted criteria. this might include existing medications, outstanding conditions a patient may have, or other pieces of information. Another very fast changing category of decision support system might be those employed by military systems, while AI may be able to speed up decision making, a decision support system needs to be fully explainable in the content of those decisions, because in military and health mistakes are not only costly but deadly as well.

As artificial intelligence continues to evolve DSS platforms will increasingly need the ability to both explain their outputs and to justify their decisions, without this critical feedback process these systems may experience errors which are propagated it's similar to the idea of a person giving an instruction and by the end of the instruction you have a completely different idea, which is unlike the original when a decision support system makes an error that error may be propagated through out the entire decision process leading to erroneous responses.

Various types of DSS

The applications of decision support systems are practically endless ranging from medical, military, financial, even cybersecurity. But again these fields have a varying degree of tolerance to errors, medical DSS it's absolutely imperative that ALL outputs contained within a decision subset be checked against actual knowledge, medical errors already cost considerable amounts of money to remediate and mitigate, but a decision support system, is the responsibility of the company or organization for the design of the system that leads to critical errors. As such actually designing a proper DSS for say military decision making, needs to not only be verified in testing but also outside of testing, this means that top military staff and personnel need to check the contents of every decision a system in this field cannot and should not be completely trusted in it's outputs.

Any critical sector which has a requirement of being absolutely correct or close to absolutely correct and is intolerant of errors should be closely monitored after deployment, and potentially risky decisions evaluated outside of the operation window for decision support. Most fields I've mentioned, cybersecurity, military, medical, financial and legal all have very high precision and accuracy requirements.

Foundations of an effective DSS

A strong foundation for a DSS is critically important for their implementation as iterated previously must be able to explain their decisions, provide an "confidence" score, and the staff operating it must also be acutely aware of these systems being vunerable to their own bias, silently fail and rigorous training and retraining. Various techniques can be used for the analysis of a DSS and many existing AI models can be used to framework them, one well known technique of a DSS is the method of restrictive k-means clustering analysis, in this we use techniques to collect "data" into a hyperplane containing the "data points" now what these data points mean might not entirely be known at this time, data science, data engineering, and machine learning engineering are all critical parts of extracting information ready for decision support systems.

Bedrock of a DSS is Data.

Just like every machine learning model data is the core driver of a DSS no data you cannot make insights into the data, this one foundational core element often glossed over by many companies who want to implement these systems, but data can be organized or it can be disorganized requiring a phase called "preprocessing" often data scientists are involved in that preprocessing they may use tools to find relationships in the underlaying data and those relationships form the basis of preprocessing the data. Here this process takes the form of both data engineering and data science. The data engineering aspect often cleans up the data making it more computationally feasible to deal with and the data-science aspect finds patterns within the data to build better and better datasets Data itself can come from many different places as stated above some organized, and some very disorganized. This data often needs to be processed before any useful insights can be drawn and then models trained on the data collected, the types of data that might be of either class are. The main source of data is usually that of databases, mostly being those from two different camps of design NoSQL and SQL depending on the application the choice of database is dependent on the type of data being stored, highly structured data, which can be incredibly useful to develop machine learning insights as it's both highly discrete and easily accessible, lending itself to be used in machine learning applications easily. Where as NoSQL databases are often good for things like relationship storage, such as those found in graph based or in the form of time series.

Sources of data

Often one core piece of data is that of organized, as this is often easily accessible to an organization, or business it means obtaining data is often best to start at databases, which is often highly organized already, meaning it needs much less work to categorized, feature engineer and configure.

Databases

Database choice is not a simple as one might think, intentionally many would assume SQL was the best option however this is a deeply flawed method of thinking, as databases have technical performance measures, these are critical. Databases that perform realtime processing or near realtime might be more suitable to using in memory methods such as those provided by Redis a database which runs in memory or hardware considerations in the form of choosing very high performance SSDs like those from the likes of KIOXIA.

• Relational Databases: Also known as row-column databases you query rows and columns based on relationships between the database entries for example SQL will use a syntax like SELECT FROM CUSTOMERS WITH PRUCHASES > 10 or something similar this relational structure lends itself well to systems like algorithms used for clustering data into groups which are effectively mostly pre-labelled.

  
  

• Document: Store these objects called documents

  that are often collections of data that might be documents or they might be collections of strings, the versatility of document databases are quite wide, and they are quite intensive to process as they contain lots of text information.

  
  

• Scientific: This is an interesting subset of many different databases and data storage mechanisms, one particular example here if you are looking for information, data science might be heavily used on for example sampling of radiowave data, for example Australias large radio telescope array processes gigabytes of data as they scan the key looking at celesetial objects and their behaviour this information is often "fairly" organized, but unique in that radio telescope data is multi-dimensional right off the bat. Some scientific databases may also be "largely text" for example the bioinformatics field is just mostly gigabytes of text data, in the form of sequences in both RNA, DNA and Protein coding. One common method is the FAST and FASTA data storage methods for these sequences. Other information from scientific databases might include particle collisions CERN uses a lot of databases to extract information about the information they collect it also contained very large amounts of logging information from detectors, to safety systems, protons as they are accelerated as dangerous and no one should be in the cooridores during a particle event.

  
  

• Time Series: As it says a time series database keeps track of time, but bare in mind a time series database can have many different types of "kept time" ranging from that which is produced by weather data, to a database that keeps track of peak traffic in a city and optimizes ambulance response time based on a small subset of "possible paths" as a solution to the traveling salesman problem reducing the time needed spent in potential traffic, they can also be used to analyze peak traffic patterns and determine what roads and sections of road could benefit from some optimization such as attempting to distribute vehicles over multiple paths to improve the load, and prevent "junction bottlenecks" within a transportation network.

  
  

• Graph Database: Information is stored in a graph this lends great use to storing relationships about customers, social media, accounts, even government systems, and decision support systems, graph database can be used to store information such as the effectiveness between each node over the graph where each "node connecting to a vertice conveys the information for a specific condition and it's effecitveness in that particular application, while also containing a "connection" to graphs that contain adverse risks or side effects. Crime databases might contain a type of graph based relationship it might also be mixed with document based databases where a document contains information at each node of the graph. We can use paths between graph nodes as relative strengths but we can actually store graph edges "paths" with a multidimensional characteristic. for example you could potentially use a numpy array to store variation information parameters which are examined at each graph node traversal.

General Data

Other source of information for a DSS is general data, now general data is a bit of a problem to deal with it must both be processed and then made into actionable insights which can be used to inform the development of a decision support system.

• Datasets: Often heavily used there is both academic and non-academic datasets that exist, academic datasets are those collated together to form the information for many "academic" fields. These datasets are not typically intended for

• Company specific information: wikis, internal websites, documents, log files, SEIM information the amount of geneal data is pretty much infinite, leveraging this data is critical to a DSS.

• Books: in this case it is very important to obtain permission and the respective purchasing rights to be permitted to train on this information, otherwise you will potentially be breaking the law opening for litigation, obtaining both written permission and licensing rights is critical, as encyclopedic information is invalauble for systems which are built on high accuracy, high precision, and reliability as their core tenants, medical DSS's can benefit significantly from being able to either be trained on the data or use RAG (As long as the results are accurate) for detecting issues in decisions within hospitals.

After data has been collected and identified it can often be transformed by ETL (Extract, Transform and Load) in doing this you can actually convert one datatype to another, for example the number of financial transactions and the country they are originating from become a useful source of data as this might be stored into a "relational database" but the time added information to how often that transaction happens and which country of origin is say a bank tracking fraudulent tranactions this might indicate on a time series database as large outgoing transfer in a short period of time. Could imply this account is compromised as the customer is in say USA or Europe but the transactions are rapid and coming from India, Pakistan, or another location within Europe that is not normally associated with previous activity of that customer.

xAI (Explainable Artificial Intelligence)

Explainable artificial intelligence is a critical component of how to build AI DSS's as all decisions made by an AI system must be "explainable" this means a human can interpret how the entire system works, and what it's outputs are, without explainability decisions from an AI DSS will become problematic as many machine learning models are blackboxes and in decision support systems the importance of making blackboxes more transparent if ultimately desired especially when AI DSS platforms begin permeating into fields of both high precision, and high accuracy like those found in the medical industry. To ensure a DSS is reliable it either must be heavily validated using rigorous and broad stress testing in the case of using a blackbox based model, or it must use explainable AI implementations at it's core to ensure the decisions made by a DSS are accurate, precise, and reliable, while offering minor deviation from perturbation noise, this may also include

Pretraining and Post Training

Depending on the application both maybe used, pretraining to ensure the system or model is accurate in it's predictions and post training to ensure it can cope for example with advesarial attacks or other machine learning based threats to the normal operation of models.

Modal Type

While many DSS will use mostly text data or text based information it might be the case multimodel decision support systems will be used, why is this the case, in some applications a DSS may need to handle multiple sources of information, and the type of modal data can affect the overall performance and reliability of a DSS. The type of model is usually one of two types, discriminator, generators and integrator. Generator: The one most people are familiar with is that of a generator with the likes of ChatGPT, Gemini, CoPilot and others allowing the creation of images or video. Discriminator: does exactly what it sounds like is separates information from the whole, be that an identification of a particular class of object, group of objects, or identification of specific data points and collecting then organizing them through datascience to build that discriminator. integrator: integrator are the real place a decision support system gets it's power oftgen including discriminator capabilities with generative, becoming and integrative system. For example many AI based platforms such as those in vehicles should be a type of integrator, single data like that found in some vehicles which are driveless lack the integrative phase where multiple pieces of information borrowing from the idea of realtime operating systems perform consensus based decisions combining the likes of rule based, with system information and well calibrated blackbox platforms that are regularly monitored and updated. Integrators also fall under the category of those used in medical situations, where information about patients is combined with system information normally databases to verify the integrity, reliability and validity of the system.

• Text: Is not just limited to document processing, but also numerical data, labelling, and categorization, again related to classification overall. Text data can also include logs, event triggers, even programming code, and documentation. But documents can be a critical part of a DSS.

  
  

• Audio: Call recordings, in call-data extraction, data extracted from video, information from music (preferrably) that which is available to creative commons or other permissible sources of information, a DSS needs ethical data. One example of DSS platforms could be used with audio data is that of updating and collecting call information directly from an active call for emergency services. Firefronts often update quickly and information may not be immediately available a highly accurate discriminator, combined with integrator can take the information from call data and convert it into NLP information to track in realtime fire front updates.

• image: Leveraging vision processing as it's core process also known as machine vision, is often used for classification, identification, object tracking, and inferencing, Image processing for multimodal platforms could include those being leveraged for example tracking illegal fishing in an area.

  
  

• Video: Processing video can be done in one of two ways, removing frames, or pieces of frames for processing, or full multidimensional processing of video data, this often requires advanced signal processing techniques including multidimensional fourier transforms and convolution, most video processing itself is really an extension on image processing to use multiple frames to inform decisions.

  
  

• Data: Is an interesting particular type of information and is just "general" in the sense it can be any kind of data collected from any kind of sensor it may not even be video, image or text it might be something much more complicated like radar, lidar, gas sensors, temperature, and it may not be "explicitly" text as text may not be the most efficient mode of storage, data might also include geospatial intelligence information.

Large Langauge Models and RAG (Retrieval Augmented Generation)

Does have the capacity to be used for DSS's IF the user of the DSS is fully aware that all outputs might potentially contain errors or issues, RAG can be be used to augment decision support platforms. In this case information from a DSS can be used to collect information, input data, such as that from simple cases like Excel spreadsheets many organizations use them for information storage about internal company project management. Large langauge models allow you to ask questions about the data within your DSS platform, in the form of a chatbot, which when combined with tool use capability becomes very powerful as a means to make decisions improve outcomes, enhanced reliability, and look for deeper insights into the data the DSS has access to. An example would be lets say we have a company with a very large and comprehensive database of knowledge based information about particular products. Example could be a network engineering company, and a customer has a question about a product multi-model processing a DSS monitors the call and looks for key information, predicting ahead of time what the customer service representative might require, such as asking about a particular model, and a DSS looks through the company knowledge base for the information a customer wants while on call taking, a lot of the drudgery out of dealing with customer service and ideally improving call experience, in this situation RAG can generate NLP requests to the wiki / knowledge base database and retrieve it as the call progressed. I think I've throughly talked in depth about DSS in this first Part I'll work on part 2 and develop the concept even further.