10/12/17: Rationally-Irrational AI, yet?..

In a recent post (http://trueeconomics.blogspot.com/2017/10/221017-robot-builders-future-its-all.html) I mused about the deep-reaching implications of the Google's AlphaZero or AlphaGo in its earliest incarnation capabilities to develop independent (of humans) systems of logic. And now we have another breakthrough in the Google's AI saga.

According to the report in the Guardian (https://www.theguardian.com/technology/2017/dec/07/alphazero-google-deepmind-ai-beats-champion-program-teaching-itself-to-play-four-hours),:

"AlphaZero, the game-playing AI created by Google sibling DeepMind, has beaten the world’s best chess-playing computer program, having taught itself how to play in under four hours. The repurposed AI, which has repeatedly beaten the world’s best Go players as AlphaGo, has been generalised so that it can now learn other games. It took just four hours to learn the rules to chess before beating the world champion chess program, Stockfish 8, in a 100-game match up."

Another quote worth considering:

"After winning 25 games of chess versus Stockfish 8 starting as white, with first-mover advantage, a further three starting with black and drawing a further 72 games, AlphaZero also learned shogi in two hours before beating the leading program Elmo in a 100-game matchup. AlphaZero won 90 games, lost eight and drew 2."

Technically, this is impressive. But the real question worth asking at this stage is whether the AI logic is capable of intuitive sensing, as opposed to relying on self-generated libraries of moves permutations. The latter is a form of linear thinking, as opposed to highly non-linear 'intuitive' logic which would be consistent with discrete 'jumping' from one logical moves tree to another based not on history of past moves, but on strategy that these moves reveal to the opponent. I don't think we have an answer to that, yet.

In my view, that is important, because as I argued some years ago in a research paper,  such 'leaps of faith' in logical systems are indicative of the basic traits of humanity, as being distinct from other forms of conscious life. In other words, can machines be rationally irrational, like humans?..

13/02/2011: What a Jeopardy champ can do in the world of finance

Here is my article along with Shanker Ramamurthy that was published last Thursday in the American Banker, discussing IBM's Watson super computer system's potential applications in the financial services industry - helping to advance industry thinking on how in the era of "big data" only advanced non-linear analytics can make sense of structured and unstructured data flows to transform it into valuable insights.

VIEWPOINT: New Computer, New Modeling Possibilities
By Shanker Ramamurthy and Constantin Gurdgiev
February 10 , 2011 - p8

Next Monday a new IBM computer system called Watson will battle two quiz-show champions in a game of Jeopardy! There is more at stake here than winning a game. The potential applications of this technology to transform the operations of industries such as health care, government and finance are enormous.

In the financial services industry, integrated risk management is an everyday struggle. Financial practitioners and supervisory and regulatory authorities must make split-second decisions using information coming from all sides: the Internet to corporate and call center channels.

The challenge is to efficiently process diverse data streams and pick out relevant data insights to apply to strategic business and regulatory decisions.

In the banking industry today, data "fuzziness" abounds. Uncertainty exists about the quality of data, assumptions and models that are being used to make judgments. This, of course, clouds the true picture of risk and biases our decision-making, often in econometrically undetectable ways.
Most banks today run risk models on a discrete and disaggregated basis while relying on often subjective assumptions. High-performance computing advances, represented by Watson's capabilities, can rectify this - by providing visibility into concentrations of risks and risk-related activities, as they happen. Simultaneously, it deploys nonlinear analytics in selecting both the statistically and operationally important scenarios.

The beauty of a nonlinear computer that "learns" is that it can analyze a complex set of implied possible scenarios and give answers to the broadest set of questions. This potentially can lead to the emergence of analytical systems that not only report on probabilistically likely events but also identify latent "Black Swan" events and even sense deeper levels of uncertainty.

For example, a legislative decision altering a specific set of financial strategies can have no impact on traditional linear models because the outcomes can be weighted by an extremely low assigned or assumed probability. But in a nonlinear world, such an outcome can still be testable as part of the selection list for reporting. More importantly, it can be made recognizable by the analytic system and, therefore, objectively reportable.

A system like Watson has the potential to get answers to incredibly difficult questions about strategic decisions, risks and market changes that can otherwise be elusive.

For example, it has the ability to create an interactive risk-pricing system using a menu of models that evolve as the system learns, detecting structural breaks in data before analysts can spot them and build them into existing programs.

Of even more significance, Watson will be able to deliver scenario analysis based not just on either event probability or expected loss/gain but also on more complex company objectives.

This can involve analyzing corporate strategy inputs, including non-quantifiable questions, alongside fully quantifiable inputs. Imagine asking a computer "How do I increase my loan book profit margin by 10%?" or "What actions can I take to strengthen my capital reserves, with minimum impact to my asset base?"

At a much deeper level, the nonlinear learning capabilities that Watson pioneers can lead to the creation of systems that are able not only to handle traditional risks and their interactions but also to evolve into systems capable of transforming deep uncertainty into explicit models. Though still some years away, this could mean an artificial intelligence able to sense Donald Rumsfeld's famous "unknown unknowns," converting them into specific models suitable for risk analysis and getting meaningful, actionable responses.

The real-time, decision-making capability that is so sought after in the financial industry will be a crucial, competitive differentiator.

As risk intensifies within interconnected global markets, the complexity and exploding volumes of data will only rise.

Shanker Ramamurthy is the general manager of banking and financial markets at IBM Corp. Dr Constantin Gurdgiev is the head of macroeconomics in the Center for Economic Analysis at the IBM Institute for Business Value.