Last week we defined Machine Learning (ML) and articulated how ML is just one of many tools under the big umbrella of AI. Just like any AI tool, ML is great at addressing certain tasks when there is ample data (such as clustering or classification), but insufficient by itself to address more complicated issues. To address these complex problems, you will require a full repertoire of AI tools. In this week’s blog, we’ll use a poker analogy to demonstrate the value of a robust approach to AI.
You’ve Got To Know When To Hold’em…
The outcome of a single hand of poker is not purely dependent on your decisions. Rather, the outcome of a hand is dependent on the decisions of all players seated at the table. Whether or not you are successful is a strategic problem.
Your name is Alice and you are seated at the table below. There are four other players at the table: Charlie, Eric, Dave, and Bob. Each player is self-interested; namely, he or she wants to win the hand and the other players to lose. The challenge here is that there is only minimal information available, e.g. the cards you’re dealt and the community cards on the table shared by all players. While you might intuit what cards another player is holding, you don’t know with absolute certainty. And without absolute certainty, you can’t be sure if the decision you are about to make is the best possible decision. You might make a great play…or you could make a terrible choice and pay dearly.
You have observed the players and their unique playing styles for some time. You have collected the most data on Charlie; you have limited data on Bob; you have no data on Dave or Eric. Because you have adequate data, you employ ML to build a behavioral model of Charlie; based upon your model, you are able to determine Charlie has a tell: every time he has a strong hand he unconsciously touches his left ear. This information should influence how you approach future interactions with him.
The player to your right, Bob, you have limited data on. You can use an AI tool like Decision Theory (DT – the study of reasoning behind individual choices) to establish a baseline for your decision-making model, with ML serving to bootstrap this model and make it more accurate based on Bob’s unique play style.
Since there is no data available to infer from either Eric or Dave, leveraging ML only offers Transfer Learning (TL – knowledge gained solving one problem applied to a different but related problem). In some situations, TL is sufficient; however, your sample size is too low – you’ve only learned from one player completely (Charlie) and partially from another player (Bob). Due to incomplete information and the ability to bluff, there are numerous styles of playing poker, so learning only one or two styles doesn’t offer a significant advantage. However, because you are generally aware of what constitutes good and bad play (a representation of another powerful AI tool, Game Theory: how outcomes are affected by multiple decision-makers via their interactions), you can use your DT model along with Probabilistic Reasoning (using the likelihood an event will occur to reason over incomplete or uncertain data) to make anticipated good decisions, despite not learning Eric or Dave’s style yet. As you learn more about Eric and Dave, you can improve your decision-making (ML) and win more hands.
That Was a Crazy Game of Poker
The poker analogy above is even more apt when you consider an AI bot co-created by Tuomas Sandholm beat four human players on the way to winning the 2016 AAAI Annual Computer Poker Competition Heads-Up No-Limit Texas Hold’em competition for the second year in a row. Libratus, the poker-playing AI co-created by Sandholm, beat Dong Kim, Jimmy Chou, Daniel McAuley, and Jason Lef, four of the world’s best players. Even so, creating bots for more than two players is still an open problem, which is why No-Limit Texas Hold’em is a game AI teams have focused on over the past few years. (Worth noting is that one of Avata Intelligence’s advisors, Vincent Conitzer, studied under Tuomas Sandholm at Carnegie Mellon University.)
We can extrapolate the poker analogy to dealing with real world issues. When we are tasked with making critical decisions – in security, for instance – and are missing data or are only presented with fragments of what we require, we can draw on a full arsenal of AI tools to conceive and execute a plan of attack. The best solution is one like Avata Intelligence’s AVA, a comprehensive AI solution that marries the right AI tools – whether they be ML, DT, PR, GT, or other capabilities – to analyze data, fill in the gaps, and provide recommendations to stay ahead of developing situations in real-time.
Next week we’ll dive deeper into the challenges the world faces and what assistance AI can offer to overcome these challenges.