Even as the Indian government is exploring all angles to ferret out the terrorists behind the serial blasts in Mumbai city on July 13, University of Maryland researchers led by a computer science professor, V S Subrahmanian, are using mathematical models and multi-player game theoretic models to make forecasts based on terrorist behaviour patterns.
"We have finished our data collection about the Indian Mujahideen (IM) and Jaish-e-Mohammed (JeM) and already derived the Stochastic Opponent Modeling Agent (SOMA) data mining and behavioral modeling engine rules. However, we have just begun to analyse SOMA rules on these terrorist outfits and so there is no paper available yet. One of the results about IM that leap out is the high probability that in years when IM has transnational links, it carries out multiple attacks on public sites -- such as the July 13 (Mumbai) bombings," noted Subrahmanian in an email response.
Pakistan-based groups like Lashkar-e-Taiba (LeT) and Jaish-e-Mohammad, he explains, have provided training for IM and even though IM attacks may not be ordered by Pakistan-based groups, it was IM's relationship with them that gave them the capacity to carry out attacks. "At the same time, there is also a high-probability of such attacks when India is undergoing internal conflicts and when it is engaging in negotiations with Pakistan. Also, when IM carries out attacks on symbolic dates (such as Wednesday which was the birthday of the Mumbai attacker), it tends to claim responsibility," he adds.
His research team developed studied five entities -– the US, India, the Pakistani military (including the Inter Services Intelligence -- ISI-- agency), the Pakistani civilian government (not including the military or ISI), and Lashkar-e-Taiba (LeT). They also looked for Nash equilibria, named after Nobel-prize winning economist John Nash, whose life was immortalised in the Oscar-winning movie 'A Beautiful Mind'. Intuitively, Nash equilibria specify situations where no entity involved in the game theoretic model can "do better" without upsetting another agency. "We did not find a single Nash equilibrium in which LeT exhibits good behavior in which the US expands financial aid to Pakistan," says Subrahmanian, adding: "This is consistent with the recent decision by the Obama administration to cut $800 million in military aid to Pakistan."
Subrahmanian is also Director of the Center for Digital International Government (CDIG), and Co-Director of the Laboratory for Computational Cultural Dynamics (LCCD) at the University of Maryland. Among other things, he has worked on the development of techniques for realtime monitoring of terror groups. Over the years, he and his team collected data reflecting about 770 variables relevant to terror groups -– this data has been collected for many terror groups operating in South Asia including LeT, JeM, SIMI, and Indian Mujahideen. The data has been collected on a monthly basis for the period of existence of these terror groups (so in LeT’s case, about 20 years, but in IM’s case, just a few).
The University's SOMA engine, then, tries to learn conditions on the environment that are good predictors of the group taking actions. They are rules of the form. When some condition is true in the environment in which the group operates, there is a probability of P that it will carry out action A with intensity I. Multiplayer game theory models, on the other hand, associate a set of actions that each player can perform. In the LeT game theory paper (accepted for publication at the 2011 European Conference on Intelligence Security Informatics and the 2011 Open Source Intelligence Conference this September), these actions are at a high level (e.g. covert action or coercive diplomacy -– but policy makers still have the option of deciding exactly what type of covert action/coercive diplomacy to use).
"For each combination of actions that the group can take, we assert a “payoff” for each group. Intuitively, if we have the 5-players in our model (LeT, Pak military, Pak civilian government, US, India), for each combination of actions these 5 players can take, we need to assert how good or bad that scenario is for each of the 5 players. This yields something called a payoff matrix, showing all possible combinations of actions, and all possible payoffs for each such scenario," says Subrahmanian.
Subrahmanian believes the SOMA model provides credible results. In a paper published in April 2008, he used probabilistic rules generated by the SOMA data mining and behavioral modeling engine to state that Hezbollah (in Lebanon) usually does not carry out transnational attacks on Israel when participating in Lebanese elections (and when certain other conditions hold). That November, the Beirut Daily Star carried an article which said that if these predictions were correct, Israel’s generals could rest easy in the early part of 2009 when these conditions were met (including Lebanese elections) – instead of investing in resources against a Hezbollah threat from Lebanon, they could use those funds for other strategic or tactical purposes. "This is exactly what happened in the first half of 2009 – virtually no Hezbollah-backed attacks on Israel. We wrote a piece on this in Foreign Policy in 2010," asserts Subrahmanian.
He concludes that "although models never capture everything and can often be wrong, the same is true of even the smartest of analysts...We strongly believe computational methods must be used in conjunction with human analysts to derive the best of both computational analytics and human subject matter expert knowledge".