Perfect decision-making requires perfect information
To make perfect decisions we require perfect information: about our environment, about own preferences, and about the preferences of others. We need to know what exists, where, and how it works. That information needs to be up-to-the-minute and real-time in nature. It needs to be accurate, relevant, and trustworthy. And at the same time we need to know what not to focus on: what factors, variables and information are irrelevant (or misleading) in any given scenario. In practice, perfect information may be unobtainable, but with IoT sensors and streaming data, we can get a lot closer than historically imaginable.
As humans, we informally model thousands of events every day, factoring in dozens or more things at once. As you read this page, you may be wondering whether or not the author sounds trustworthy, whether the information is of interest to you, and whether or not to continue reading. To do so, your subconscious is mentally modeling the world around it: processing information on the fly, and producing judgements. With a little conscious effort, we can often drill down into our decisions to try to figure out why we believe certain things to be true, or feel a certain way, but much like neural nets and other deep learning, explainability in particularly complex multi-factor environments can often prove elusive.
The power of modeling is profound. The outputs of one model can be taken as the input of another. And so hundreds of models chained together effectively enabling ‘decision-making’ under any set of conditions.
A brief history of agent-based modeling
Agent-based models (ABMs) have existed within academia since at least the early 1970s. The very first ABMs were produced using coins and graph paper, with agents of different types (represented on a grid by different color coins) interacting with each other and their environment on the basis of simple rules. These agent-based models, known as ‘cellular automata’, even limited to small number of agents, proved capable of demonstrating the emergence of complex phenomena from very simple rules. The best known such model is perhaps John Conway’s Game of Life model.
Through the power of modern distributed computing power, and the open-source HASH platform, it is possible to build massively detailed agent-based models capable of capturing far more emergent phenomena.
Most ABMs in the 70s (and still today) are heavily abstracted, both due to the costs of collating quality data and mapping them to agent schemas, as well as the time-sink of specifying detailed rules of interaction between agents and their environment. HASH aims to solve this problem in a number of ways:
- Through the hIndex data and models repository:
- making pre-normalised and schema-mapped data quickly available for inclusion in ABMs; and
- enabling the free sharing and quick forking, as well as purchase and sale of models a possibility.
- Through our open-source hCore simulation engine:
- supporting the incorporation of real-time streaming data sources; and
- enabling the embedding of detailed specialist models and their outputs within larger system ‘meta models’.
- Through our hCloud:
- enabling the massively-distributed computation of simulations of an unlimited size.
The future of decision-making
We imagine a future in which no major decision is made without simulating its consequences, and the potential appropriateness of alternative approaches.
ABMs are highly explainable and scrutinable, relative to many machine learning approaches, and through HASH’s hCore graphical user interface subject to review by domain experts as well as programmers and data scientists.
Today, ABMs can be used to support decision makers by helping:
- provide understanding into emergent phenomena that are found within complex systems;
- uncover unknown risks that result from the interdependence and interactions of agents;
- test specific hypotheses against each other to determine preferable strategies for navigating a system;
- automatically identify the optimal balance of resources to deploy through Monte Carlo simulations.
Whilst supporting human decision-makers today, we imagine a future in which battle-tested ABMs are able to handle a majority of ordinary business decision-making in a fully automated fashion.