How Evolutionary AI will inform future business decisions
Creating an AI-Powered Enterprise
Evolution and decision-making are not immediately linked in our minds; however, as it turns out, algorithms inspired by biological evolution are the key to augmenting decision-making in a wide variety of business use-cases.
But let’s start with the problem statement. My team and I are continually engaged in conversations with enterprises from various industries about their expectations for artificial intelligence. Often, we learn they’re seeking better ways to model the data that flows through their systems.
We hear questions as to whether AI can:
Better predict crop yields.
Improve diagnosis accuracy by classifying MRI images.
Predict the risk of insuring properties.
Rate the likelihood of a transaction being fraudulent.
Identify user segments likely to purchase a premium subscription.
These questions are all about using AI to produce more insights. If the AI systems above were actually already built and deployed, businesses would be able to:
Buy the correct amount of seed.
Apply the most effective treatment.
Decide whether to insure the property.
Reduce fraud through preemptive action.
Optimize marketing spend and increase revenues.
Limits of Decision-Making With Current AI
But while today’s AI applications handle analytics well, the resulting decision-making is still done manually or by using rigid algorithms designed to optimize one or more outcomes. Insights are provided, but the choice of which action to take is not always obvious.
Let’s say an executive has determined, through the use of advanced modeling and projections, that the business is at risk of missing its revenue targets in six months. What should be done? Cut costs? Sell at a discount or renegotiate supplier prices? Retarget sales? Or maybe the right choice is to invest in growth or efficiency. Which combination of these choices is best? Given one of these choices, what is the best way to implement it?
Simple extensions to reporting, such as predictive analytics, produce coarse grain advice. Even if your system determines your costs are high, it’s still up to you to decide what to do about it. Businesses want to maximize revenue, maximize sales and minimize cost and risk, but these objectives are often not aligned, and focusing on one usually comes at the expense of another. Reducing costs by too much can mean losing valuable talent, and over-emphasis on sales can lead to unsustainably low prices. Neither of those options is viable. The best decisions are made when we correctly balance these factors and objectives.
Further, current AI systems have a hard time effectively dealing with changing conditions. Today, decision-making models often become obsolete more quickly than we expect, with rapid changes in habits, culture, fashion, the economy, technology disruption, and legal and regulatory changes. Models based on historical data also cannot withstand what are often called “black swan” events.
Surely, one might think, AI systems that can beat human world champions in chess or Go can perform advanced decision-making? Surely if a system can beat humans playing video games like Flappy Bird, it should be able to help with business decision-making?
But the answer is no. Even systems that learn approximately a million permutations are still not agile enough to respond to unexpected states that have not been captured in the training set. They fail miserably if game rules are changed.
Evolutionary Computation Is Shifting the Paradigm
How can we overcome these limitations and build an effective decision augmentation system? We will need to program machines to make more human-like decisions. But, how is this done?
With human decision-making, every decision is based on a context: on which actions are taken, with the goal to achieve the best outcomes. As humans, we draw conclusions by continuously forming and updating mental models of the world, creating hypothetical contexts, and then mentally forecasting the implications of different actions based on historical knowledge of previous outcomes and an understanding of current conditions and how they might change over time. We then apply our chosen strategy, observe the results, and then modify and improve our mental models of the real world.
Figure 1. With prescriptive continuous learning and optimization, context, action and outcome (CAO) data is aggregated from the real world.
Part of this process is familiar to us: Building a mental model of the real world is analogous to model building—something today’s AI is quite good at. We refer to this model in Figure 1 as a predictor, and its job is to take historical decision data, along with the resulting outcomes, and build a model. This model will not always be accurate, especially when faced with unfamiliar context/actions, but it does provide a degree of confidence in its projections.
These outcomes need not be—and often are not—immediate. We are often faced with outcomes that lag our actions; as long as the lag is reasonable and the resulting outcome can be associated with the past context/action, however, a predictor model can be built using standard data science.
Given a predictor model, how can we build a prescriptor that suggests actions given a new context, such that the outcomes are optimized? We need a way to come up with a variety of strategies and try them out against our predictor, which would act as a surrogate for the real world.
How Evolutionary AI Works
This is made possible through the use of an AI technology referred to as Evolutionary Computation. Evolutionary Computation is a population-based approach. This means that rather than modifying a single solution until it is in an acceptable state (e.g., back-propagating deep networks), it:
Generates a population of candidate solutions (initially randomly).
Calculates their “fitness.”
Gets rid of the less fit candidates.
Regenerates new candidate solutions by either randomly tweaking the fitter ones or treating fitter ones as parents and borrowing traits from them and generating new candidates.
In other words, we use Evolutionary AI to generate a population of models that map a given context to proposed actions, and we rate them based on how well they do on the outcomes predicted by the predictor.
The process of training a predictor model and evolving a prescriptor model is called Evolutionary Surrogate-assisted Prescription (ESP), and it is available as Cognizant’s AI Optimization Service through our proprietary Evolutionary AI platform.
A model of this approach can be seen in Figure 2. The background to this chart is the ground truth, or the real world. Each decision is represented by a dot, and the optimum decision, in this example case, would always be on the y = 0 line through the middle.
In Figure 3, the AI Optimization Service loop is animated. As a decision (represented by a white dot) is made in the real world, a predictor is created as a back-propagated neural network. A prescriptor is then evolved against it, and the prescriptions from the prescriptor are applied in random contexts in the real world. After a few data points, the decisions start to align neatly with the optimum line shown across the middle. Note that the predictor’s model of the world, which is the background, starts looking more and more like the ground truth from the prior figure.
Our AI Optimization Service has been used successfully to discover recipes for growing plants in controlled environments, where the historical decision data is sparse and few, and the outcome lags by six to eight weeks—the time it takes for basil to grow from seedling to full plant.
Figure 2. Ground Truth
Figure 3. Optimization in Action
Evolutionary AI in the Real World
Examples of the AI Optimization Service’s usage in business decision-making are abundant. In a company like Cognizant, an important decision cycle has to do with responding to software development and testing requests for proposal. The RFP often includes ample contextual information that informs decisions on the best mix of developers, tools and time to budget for each given RFP. The outcomes being optimized are quality and speed within a reasonable cost.
In an insurance company, the most important and impactful decision cycle is underwriting, which comprises information about what is being insured—the context; the decision of whether or not to insure, with what kind of policy and at what premium—the actions; and with the goal to win the customer’s business, while minimizing the loss ratio measured a few months after the decision—the outcomes.
To compare our AI Optimization Service with reinforcement learning systems, we applied it to the Flappy Bird game.
In the first few iterations, the system started to form a model of the Flappy-Bird world, and given the sparsity of the data, its models are almost hallucinatory in the example on the left. But things started improving. After a small fraction of actual games combined with deep reinforcement learning approaches, we were able to play the game at human levels. This is shown in our example below on the right.
Example 1: In the initial iterations, the bird randomly crashes into the columns and the predictor is quite inaccurate regarding the game and outcomes.
Example 2: After 200 iterations, the bird masterfully navigates all levels of the game.
The Future of Business Decision-Making
To summarize, AI enabling decision-making through our AI Optimization Service allows multiple objectives to be optimized simultaneously. This is achieved via evolutionary computation’s inherent ability to pursue multiple objectives due to it being population-based. This means that the system can reveal “the art of the possible” when it comes to the different outcomes. Within the AI Optimization Service’s executive dashboard, executives can set a target for a certain outcome to see not just whether that target is achievable but also what the best achievable targets for the other outcomes can be.
Note that the decisions are augmented at a granular level, every time a decision is made; however, the outcome goals are set at an aggregate level. Unlike coarse-grain systems, such as predictive analytics, this makes the decision-making highly contextual and, therefore, much more effective.
Key advantages and characteristics of the AI Optimization Service include:
It’s a principled way of augmenting business decisions. The system prescribes decisions based on a model of past conclusions, reducing reliance on intuition and guesswork.
It doesn’t need inordinate amounts of data to start with. Once bootstrapped with historical decision data, the system itself helps decide what new decision data points to collect, making this an AI-driven data collection system.
It’s robust and resilient to changes in the world.The system keeps tabs on how well the predictor from the last iteration is doing compared with decision data collected from the real world in the current iteration. This allows it to determine the relevance of older historical data, improving the manner by which newer predictors are trained.
We believe the future of enterprise AI - enablement lies with AI-augmented decision-making. A business with a hierarchy of AI-enabled decision-makers will draw more accurate conclusions, more quickly, and will achieve much better outcomes. As it turns out, evolutionary computation is indeed the key to creating effective AI-enabled decision - augmentation systems.
Babak Hodjat is VP of Evolutionary AI at Cognizant, and former co-founder and CEO of Sentient. He is responsible for the core technology behind the world’s largest distributed artificial intelligence system. Babak was also the founder of the world's first AI-driven hedge fund, Sentient Investment Management. Babak is a serial entrepreneur, having started a number of Silicon Valley companies as main inventor and technologist.
Prior to co-founding Sentient, Babak was senior director of engineering at Sybase iAnywhere, where he led mobile solutions engineering. Prior to Sybase, Babak was co-founder, CTO and board member of Dejima Inc. Babak is the primary inventor of Dejima’s patented, agent-oriented technology applied to intelligent interfaces for mobile and enterprise computing—the technology behind Apple’s Siri.
Babak is a published scholar in the fields of Artificial Life, Agent-Oriented Software Engineering and Distributed Artificial Intelligence, and has 31 granted or pending patents to his name. He is an expert in numerous fields of AI, including natural language processing, machine learning, genetic algorithms and distributed AI and has founded multiple companies in these areas. Babak holds a doctorate in machine intelligence from Kyushu University, in Fukuoka, Japan.