SECCION IV: DATOS SOCIODEMOGRÁFICOS
5.4.1. Metodologías empleadas para el análisis de los datos
We briefly discuss a few possible directions for future work.
Measuring trustworthiness of the provider. In this thesis, we have presented semantics and algorithms for the provider to adhere to its probabilistic commitment.
A follow on and related problem is how the recipient can measure the provider’s trust-worthiness, in order to decide whether it should trust the provider and agree on the commitment. In the decentralized setting, as we have discussed in the problem of com-mitment formulation, the recipient does not have full knowledge about the provider’s environment and/or policy, thus making it a challenging problem to precisely assess
the probability of the commitment being realized. As a feasible approach, either communicating directly about the provider’s environment and/or policy, or about the provider’s historical interactions with its environment, or both, will facilitate the recipient’s assessment. If the recipient can effectively measure the provider’s trust-worthiness, we can ask how the provider can earn trust with minimum communication with the recipient and/or interactions with the environment.
Improving the recipient’s interpretation of maintenance commitments. In Chapter V, we have supported the claim that the recipient’s interpretation of mainte-nance commitments is harder by studying several strategies for creating the approxi-mate influence. A natural question to ask is whether there exists such an approxiapproxi-mate influence for maintenance, other than the ones we have studied, that we can prove has a lower bound on its suboptimality (similar to the one in Theorem V.1 for achieve-ment), and/or we can empirically show induces low suboptimality. If the answer is negative or it is expensive for the recipient to create such an approximate influence, then we might need to rethink how we represent maintenance commitments for multi-agent coordination. For achievement, the customarily terse commitment abstraction gives the provider a lot of flexibility by only constraining it to meet the probability at the commitment time and so it can unilaterally change its influences before then. In many cases, the gain in flexibility for the provider can be worth the relatively small value loss to the recipient. However, for maintenance, as it is difficult to find an ef-fective approximate influence, the potential for the recipient to lose more value could mean that the provider should commit to a more detailed specification—the loss of flexibility for the provider in this case is warranted because the recipient makes much better decisions. Potential future work can better understand such tradeoffs in us-ing maintenance commitments, allowus-ing the community to apply commitment-based coordination to domains involving both achievement and maintenance.
Efficient formulation of cooperative maintenance commitments. In Chap-ter VI, we have developed algorithms that efficiently formulate cooperative commit-ments for achievement by exploiting the structural properties of the commitment value functions. A natural question is whether these properties still apply to main-tenance commitments, so that we can develop similar algorithms for efficient formu-lation of cooperative maintenance commitments. The proofs for the properties of the provider’s commitment value function are agnostic about the commitment type, and thus can still apply to maintenance. For the recipient, its commitment value
function for achievement hinges on the minimal enablement duration influence where u− probabilistically toggles to u+ at the latest time step by the commitment time.
Thus, the proofs of the structural properties of the recipient’s commitment value for achievement cannot straightforwardly apply to maintenance. Moreover, as we have discussed, it remains an open question what approximate influence is best to use to compute the recipient’s commitment value for maintenance in the first place.
Beyond binary commitment features. For the recipient’s interpretation (Chap-ter V) and cooperative commitment formulation (Chap(Chap-ter VI), this thesis has solved these problems for the scenario where the commitment feature is binary, involving two types of commitment that toggle the feature in opposite directions. The provider’s modelling of a probabilistic commitment is nearly identical for the the two types of commitment and can be easily extended beyond the binary commitment feature.
However, future work is needed to better understand how the recipient should in-terpret and utilize a probabilistic commitment for which the commitment feature is more complicated than binary and how the two agents can efficiently formulate such a commitment for coordination.
Communication during commitment execution. We have focused on the sce-nario in which the communication between the agents is only allowed during commit-ment formulation, but not allowed during the commitcommit-ment execution phase (including the provider’s adherence and the recipient’s interpretation). We could relax this re-striction by allowing (limited) communication during execution, and a number of interesting questions could arise subsequently. Such a relaxation could lead to the problem of how the agents can best exploit the limited communication. For example, if the provider is allowed to inform the recipient, for a limited number of time steps during execution, of the probability of realizing the commitment from the current time step (e.g., the probability for the iterative lookahead in CCIL), how should the provider wisely decide when to inform the recipient? Such a relaxation could also lead to the problem of multi-commitment formulation. For example, if an achieve-ment commitachieve-ment ends up being unrealized by the commitachieve-ment time, what if we allow the agents to formulate a second commitment for the execution after the com-mitment time? Moreover, once an achievement comcom-mitment is realized, the agents can start formulating a maintenance commitment for the subsequent execution about the precondition that was just enabled, and thus such a relaxation encourages us to develop a unified framework for both achievement and maintenance.
Scaling to more agents and commitments. Throughout this thesis, we are con-cerned with a single commitment between two agents, with one agent fixed as the provider and the other as the recipient. Much future work is needed for handling scenarios where there can be more than two agents for coordination using multi-ple commitments. The provider might make a commitment to multimulti-ple recipients.
Instead of a single commitment, agents might need to coordinate with a chain of commitments that are temporally correlated. Mutual and cyclic commitments can exist, where an agent can shift from being a provider to being a recipient over time, or even can be both a provider and a recipient at the same time. These interesting scenarios naturally exist in multiagent coordination, and extending the work accom-plished in this thesis to these scenarios requires scaling the problem formulations and solution methods to multiple agents and commitments.