Machines at Play
What the success of artificial intelligence in gaming can teach us about where to apply it in the enterprise
Mackey is a Partner at OpenView, the expansion stage venture capital firm.
DeepMind’s AlphaGo caught the attention of the world in 2015 when it became the first software program to beat a human professional Go player. While the breakthrough that led to this came three years earlier in 2012, AlphaGo’s victory was seen by many as the next symbolic step in artificial intelligence’s advance toward human capabilities since Deep Blue’s defeat of Gary Kasparov in 1997. Earlier this year, AlphaGo defeated the world’s number one player 3–0 and retired with the defeated champion calling it a “god of Go.”
Since then, there have been an increasing number of examples of intelligent systems beating human professionals at a wide variety of games, most notably Elon Musk’s OpenAI defeating a professional player of Dota 2 — one of the most popular online games in the world. So what is it about games that makes contemporary machine intelligence approaches so successful at transcending human abilities?
The short answer is that games are well-defined. Specifically, they have a set of rules that govern the current state of the game, actions that can be taken to change that state, and an objective function that defines the end goal — winning. Taken together these three characteristics create a knowable, consistent environment where an intelligent system can play millions of matches to can accumulate more experience than any human. If a problem exhibits these three characteristics, it’s much more likely to be a candidate for the application of machine intelligence today.
Objective Function
What is good? In many instances, the objective is clear. Do something faster, more accurately, or at lower cost. However, there are many endeavors where quality isn’t quantitatively measured. A classic example would be attempting to write a masterpiece. The process of writing has a clear state, the words, spacing, and punctuation on a set of pages. There is also a straightforward way to change that state, add or delete characters or spaces on those pages. Yet, what makes something a masterpiece is not well-defined.
Another way to think of this is whether the goal is objectively or subjectively defined, whether ultimately the actions being taken should be measured by an objective function or subjective function. If simple knowledge of the relative state determines quality, the task could be a good candidate for intelligent software.
Known Inputs
Are the complete set of inputs observable? If not, the predictive variables may not be captured by the machine learning process. An example of an activity with a clear objective function and known state, but an incomplete set of inputs, is trading a public security. The price (state) is known and the objective function (making a profit) is clear, but the full set of drivers underlying a change in that price may not be observable. In fact, some of these factors are categorized as insider information, and if observed, are illegal to take action against.
Solving this information gap is why hedge funds are so hungry for data. The closer they can get to the machine intelligence triumvirate of a clear objective function, known inputs, and clear state, the stronger their competitive advantage. Similarly, problems well suited for intelligent software solutions have an observable set of known inputs.
Clear State
Establishing a clear state is often the first step when approaching a problem. Whether in ancient celestial navigation or simply trying to follow recipe, it’s hard to know where you’re going or what to do next unless you already know where you are.
An example of a problem with a clear objective function and known inputs, but where the state is not clear, is finding the exit to a randomly generated maze from a random starting point within it. The objective function is reducing the number of steps to exit and input is taking a step, but after that step is taken it’s hard to know if you’ve made progress. Unlike missing an objective function or having unclear inputs, neither humans nor software are adept at solving problems without a clear sense of state, although software does have the advantage of a perfect memory.
Can we relax these constraints?
If the problem you’re looking to solve with machine intelligence has a clear objective function, a knowable set of inputs to change state, and a clearly defined state at all times, it’s an ideal candidate for machine intelligence. If it lacks an objective function, pairing intelligent software with a human to apply judgement is a good bet. Alternatively, if the inputs are unknown, amassing more data about the problem could lead to a viable approach. However, if the state isn’t clearly defined, it will be hard to make material progress — as even humans are typically lost.
At OpenView we’re actively looking to partner with the next generation of intelligent software companies. In other words, if you can reduce the problem you’re trying to solve to a game — I’d love to see how your software plays.
