AI wins, as smart machines develop a ‘poker face’

IoB Insiders In poker, artificial intelligence has proved its ability to work with uncertain situations and incomplete information, with huge implications for the insurance sector, writes Andy Yeoman, CEO at Concirrus.

In the past few years, one of the areas that has seen the biggest advances in software sophistication is gaming. Put simply, games are a good way to test and improve artificial intelligence (AI) algorithms, as they provide a secure environment with clear boundaries in which software can learn rules and develop new strategies.

For an example, we need look no further than AlphaGo’s victory against Lee Sedol in 2016 in the Google DeepMind Challenge Match. Here, the AlphaGo computer program beat Sedol 4-1 in a five-match game of ancient Chinese board game, Go. Or you may remember IBM’s Watson beating human competitors in a mock-up of the TV quiz show Jeopardy! back in 2011.

Poker, however, offers new challenges and new opportunities to push AI to its limits. Earlier this year, an AI program called Libratus managed to beat four of the world’s best poker players at no-limit Texas Hold’em, winning $1.7 million in (fake) chips.

This was rightly recognized as a big milestone for AI, because it showcased the ability of algorithms to handle uncertainty and situations of incomplete information. In no-limit Texas Hold’em, players have to deal with bluffs and counter-bluffs, taking into account deception strategies from other players, as well as making judgement calls on hidden cards. That the software was able to navigate these uncertainties and still win convincingly is remarkable.

“Learn on your own”

How did the software become so good? Having been taught the basic rules of the game, Libratus proceeded to learn new strategies by running its algorithms across millions of hands of cards. As Noam Brown, the PHD student at Carnegie Mellon University who built Libratus with his professor Tuomas Sandholm, has put it, “We didn’t tell Libratus how to play poker. We gave it the rules of poker and said ‘learn on your own.'”

The strategies it came up with were both successful and unexpected. In particular, its human opponents struggled to cope with Libratus’ aggressive tactics, making huge bets for relatively small prize pots.

But what does a poker-playing AI program have to do with insurtech and insurance? Well, as Sandholm pointed out, the potential applications for this type of software are huge, “because most real-world situations are games of incomplete information.”

These include business negotiations, consumer sales tactics, cyber security strategies and medical treatment plans, for example. In short, with the capability to make intelligent decisions in situations of incomplete information and uncertainty, AI can be applied to a range of business environments.

Insurance is one of those environments, where staff are constantly making judgement calls, whether they are underwriters deciding on assets to underwrite or risk management personnel balancing their organisation’s exposure.

AI takes hold

The financial services sector has already seen AI get a foothold. At the end of last year, the world’s largest hedge fund, Bridgewater Associates, announced that it was building software to automate the day-to-day management of the firm. Its goal is to automate three-quarters of all management decisions within five years.

Exactly how this will play out in other companies and organisations remains to be seen and a debate is taking place over the exact nature of AI’s role and its relationship with human colleagues.

But for any organisation that is data-driven (or wants to become more data-driven), algorithmic technology represents a logical next step. Why waste human resources ploughing through complex data, such as historical claims information, when a computer can do it in seconds and make a more accurate decision? This frees up human resources for creative, qualitative roles.

Our view of the future of insurance is that insurers will have to become data-driven, making use of connected technologies such as the IoT in order to survive. Those that don’t are likely to become extinct within the next 5 years.

But spread betting on multiple new technologies to see what’s useful is unlikely to produce solid results. Rather, organisations need to find technological solutions to specific business challenges. AI is likely to feature heavily as a solution to a number of business challenges, helping to make sense of, and utilize, huge volumes of information provided by connected technology. This is what we’re addressing, and what we’re working with our customers to achieve.

Utilizing data is not easy however, and insights need to be implemented as business rules in order to change the way an organisation works. Algorithmic technology enables us to do this, with AI software automating decisions in a way that is both efficient and intelligent.

If an insurer wants to know the dollar value of its total exposure within the predicted storm path of a hurricane, to give a topical example, machine learning algorithms (such as those within our web app Quest) can do this for them. The calculations may include mobile assets, such as ships, which are currently difficult for insurers to keep track of. Other use cases may involve policing the boundaries of new, connected, policies. In short, this technology enables the organisation to adapt its offering and work with a different business model.

The need is clearly there, and insurers who take the first steps towards a connected future are likely to gain an advantage within the market, suffering fewer losses and making higher profits. Imagining that future may be difficult, but, right now, insurers just need to start the journey.