A summary of the 5 minute presentations by the four panelists: Milind Tambe: It is clear that the DCR community is growing and is robustly moving ahead, with new problems being tackled, new attendees to the workshop and publications from different research groups. Although we want the community to succeed and continue to grow, we should be mindful of the direction(s) that the community is heading towards. To aspire for longevity, the work of the community should have societal impact or in other words, should be inspired by real world applications. There are often discussions regarding the correct way to extend a model and the correct way to evaluate algorithms. For example, which of the many metrics - cycles, NCCCs, CBRs or simulated runtimes - should be used? If these models are extended to solve real world problems, and are evaluated in real world metrics, these issues can be resolved. In fact, solving a real world problem might bring up an entirely different issue that was not considered. There is a concern that if problems are artificially created, we could be using the wrong metric. However, this is just a concern that should be brought up. Not everyone need to work on real world applications, it is also important to develop theoretical work. However, we should be mindful and be concretely motivated by real applications. To the new and budding researchers, one should aim to have their work creating ripples in the future. A research career should not be a random walk of multiple papers. ---------------------- Makoto Yokoo: Makoto's presentation is primary on applications that require privacy and is based on his work titled "Secure distributed constraint satisfaction: reaching agreement without revealing private information" with Koutarou Suzuki and Katsutoshi Hirayama. The major motivation for using the DCR model is privacy/security, such as in meeting scheduling problems. However, in existing algorithms, some private information is revealed. To successfully prevent such information leakage, one has to utilize various information security/cryptography techniques. One possible approach is to use Homomorphic encryption, such as the El Gamal encryption, which permutes constraint matrices to create an equivalent problem. The agents can then cooperate to solve the equivalent problem, and the solution is decrypted such that the decrypted solution is a solution to the original problem. In such a case, full privacy is maintained. Possible directions for DCR researchers who would like to work on privacy include the use of third party servers, which can be passive adversaries, but are willing to perform computation tasks. Alternatively, the use of secure dynamic programming (Suzuki et al. Financial Cryptograph 02, AAMAS 02) can be applied to DCR algorithms. ---------------------- Boi Faltings: Boi presented the different applications that his group developed DCR algorithms for over the years. He started with applying DCR algorithms in the coordination of projects for CERN. Unfortunately, it was not a ready in time and CERN ended up using a centralized algorithm - ILOG. However, that work spawned a company called SyncTec that is still operating today. The next application that his group looked at was in the allocation of slots in airports to airline companies. It seemed like a good application since airline companies are entities that would prefer to keep their information private, yet must collaborate with each other to decide how the slots are allocated. Existing allocation methods are inefficient as many slots that are allocated remain unused. However, it was very difficult to change the status quo and it is not politically safe to do so. Next, his group looked at smart power grids that are able to turn power on and off on-demand. This application looked promising since a distributed algorithm will avoid single points of failures and is thus more resilient. This work is published in AAMAS 09 using real data provided by a Danish research group from Copenhagen. However, convincing the powers that be to adopt this new framework continues to be a challenge. Boi presented two other applications - sensing robots and truck task coordination - before concluding that one of the main lessons learned is that it is very difficult to get a consensus on using innovations. However, we should not despair if our innovations are not immediately recognized. After all, it is often that several innovations are needed before a breakthrough is made. Although it is often that only the last innovation gets all the credit, we should still be proud of providing the building blocks to such breakthroughs. ---------------------- Amnon Meisels: Many applications are not entirely cooperative as is the common assumption in most DCR models. Thus, we should extend DCR models to consider self-interested agents. Amnon suggested three ideas to do so. The first idea is the use of asymmetric DCOPs that include personal gains in their models of the constraint costs/utilities. Thus, optimizing the global cost/utility will take into account the personal gains as well. The second idea is to alter the objective function to include a social component of the objective function. This idea is non-trivial since the social objective functions for each agent might differ from one another and in some cases, might contradict each other. Thus, a uniform measure of personal gain using complex objective functions may require algorithmic work. The third idea is to use cooperation games that provide incentive or a rational reason for the agents to play the game cooperatively, although they might be self-interested. Work in this area can then borrow insights from researchers working in mechanism designs in fields such as negotiations and auction. ---------------------- Discussion between panelists, moderated by Roie Zivan. Topics include 1. What kind of applications can be applied by the standard model? 2. Examples of applications which do not? 3. What is required from an extension to the standard model (formal definition, examples etc.)? 4. Evaluation methods for extended models, measurements, simulators, etc? Boi: Finding applications in which DCR algorithms can be used is often not the problem. The main problem is often convincing others that DCR algorithms is a good idea! We need problems whose obstacles are primarily technical ones and not social ones. Amnon: Finding good solutions to the distributed meeting scheduling has not been too successful thus far. Yet, it is a good application for DCR models and there are many tools, such as Google calendar and the Android phones, that can help spur the implementation of DCR algorithms. The community is challenged to develop an application that models both self-interested and cooperative agents, and mimics real humans in a realistic way, such that Google will buy it! Milind: Instead of engaging in a debate about centralized vs. distributed, which always creates a big problem in what I have seen of these debates (because we are fighting ghosts), I would recommend just building an application(s) that works using DCR techniques even if there is room for such debate. If we just build something, and it works, then it will be up to others to prove that they could do something better centralized. Makoto: Although it is tempting to model all the various dimensions that exist in real world applications, we should keep in mind that we want to keep the model clean. Boi: We are largely in academia and it is often that the situation is similar for other subfields of artificial intelligence. For example, he was trying to find an application for each major idea in a textbook that he is writing, and he realized that there is often a mismatch between relevance in practice and in academia. For example, greedy algorithms and rule-based programming have large impacts in the real world and the impact of planning algorithms is small in comparison. Yet, there is a whole community and a conference dedicated to planning! Since different applications require different assumptions, a lot of adjustments need to be made before an algorithm can be applied to the application. Thus, we should study fundamental properties of the DCR algorithms such that they can be applied to a variety of applications. And to do so, we will need to keep the models clean and simple.