Martijn C. Schut

The Theory and Practice of Intention Reconsideration

One of the key problems in the design of belief-desire-intention (BDI) agents is that of finding an appropriate policy for intention reconsideration. Crudely, the idea is that at any given time, an agent will have a number of intentions, relating to states of affairs that the agent has committed to bring about. An agent chooses plans that are appropriate for bringing about these intentions; if a particular plan for a given intention fails, then the agent will typically replan, to find an alternative course of action for this intention. However, a rational agents intentions will not be static. From time-to-time, it makes sense for such an agent to reconsider its intentions, for example when the intention is doomed never to be realized, or else when the agent would simply profit from adopting another, more fruitful goal. This paper presents a detailed investigation of the properties of intention reconsideration. The work builds on the foundational work of Kinny and Georgeff, who investigated the properties of various intention reconsideration strategies in environments that were to varying degrees dynamic, i.e. subject to unanticipated change. The present paper broadly falls into two distinct parts. In the first part, the authors extend work of Kinny and Georgeff, by investigating the properties of intention reconsideration strategies in environments that are also to varying degrees (in)accessible and (non-)deterministic. They then investigate two different models of intention reconsideration. In the first model, intention reconsideration is modelled as a process of discrete deliberation scheduling: intention reconsideration is modelled as an action that may be performed by an agent, and so lends itself to analysis in terms of conventional decision theoretic models of optimal action. In the second, intention reconsideration is modelled as a partially observable Markov decision process (POMDP): solving the POMDP means finding an optimal intention reconsideration policy.

The control of reasoning in resource-bounded agents

Autonomous agents are systems capable of autonomous decision-making in real-time environments. Computation is a valuable resource for such decision-making, and yet the amount of computation that an autonomous agent may carry out will be limited. It follows that an agent must be equipped with a mechanism that enables it to make the best possible use of the computational resources at its disposal. In this paper we review three approaches to the control of computation in resource-bounded agents. In addition to a detailed description of each framework, this paper compares and contrasts the approaches, and lists the advantages and disadvantages of each.

Intention reconsideration in complex environments

One of the key problems in the design of belief-desire-intention (BDI) agents is that of finding an appropriate policy for intention reconsideration. In previous work, Kinny and Georgeff investigated the effectiveness of several such reconsideration policies, and demonstrated that in general, there is no one best approach different environments demand different intention reconsideration strategies. In this paper, we further investigate the relationship between the effectiveness of an agent and its intention reconsideration policy in different environments. W e empirically evaluate the performance of different reconsideration strategies in environments that are to varying degrees dynamic, inaccessible, and nondeterministic. In addition to our empirical results, we are able to give preliminary analytical results to explain some of our findings.

Modeling centralized organization of organizational change

Organizations change with the dynamics of the world. To enable organizations to change, certain structures and capabilities are needed. As all processes, a change process has an organization of its own. In this paper it is shown how within a formal organization modeling approach also organizational change processes can be modeled. A generic organization model (covering both organization structure and behavior) for organizational change is presented and formally evaluated for a case study. This model takes into account different phases in a change process considered in Organization Theory literature, such as unfreezing, movement and refreezing. Moreover, at the level of individuals, the internal beliefs and their changes are incorporated in the model. In addition, an internal mental model for (reflective) reasoning about expected role behavior is included in the organization model.

New Ways to Calibrate Evolutionary Algorithms

The issue of setting the values of various parameters of an evolutionary algorithm (EA) is crucial for good performance. One way to do it is by controlling EA parameters on-the-fly, which can be done in various ways and for various parameters. We briefly review these options in general and present the findings of a literature search and some statistics about themost popular options. Thereafter, we provide three case studies indicating a high potential for uncommon variants. In particular, we recommend focusing on parameters regulating selection and population size, rather than those concerning crossover and mutation. On the technical side, the case study on adjusting tournament size shows by example that global parameters can also be selfadapted, and that heuristic adaptation and pure self-adaptation can be successfully combined into a hybrid of the two.

Principles of intention reconsideration

We present a framework that enables a belief-desire-intention (\acro{bdi}) agent to dynamically choose its intention reconsideration policy in order to perform optimally in accordance with the current state of the environment. Our framework integrates an abstract \acro{bdi} agent architecture with the decision theoretic model for discrete deliberation scheduling of Russell and Wefald. As intention reconsideration determines an agents commitment to its plans, this work increases the level of autonomy in agents, as it pushes the choice of commitment level from design-time to run-time. This makes it possible for an agent to operate effectively in dynamic and open environments, whose behaviour is not known at design time. Following a precise formal definition of the framework, we present an empirical analysis that evaluates the run-time policy in comparison with design-time policies. We show that an agent utilising our framework outperforms agents with fixed policies.

Reinforcement learning for online control of evolutionary algorithms

The research reported in this paper is concerned with assessing the usefulness of reinforcment learning (RL) for on-line calibration of parameters in evolutionary algorithms (EA). We are running an RL procedure and the EA simultaneously and the RL is changing the EA parameters on-the-fly. We evaluate this approach experimentally on a range of fitness landscapes with varying degrees of ruggedness. The results show that EA calibrated by the RL-based approach outperforms a benchmark EA.

Is self-adaptation of selection pressure and population size possible?: a case study

In this paper we seek an answer to the following question: Is it possible and rewarding to self-adapt parameters regarding selection and population size in an evolutionary algorithm? The motivation comes from the observation that the majority of the existing EC literature is concerned with (self-)adaptation of variation operators, while there are indications that (self-)adapting selection operators or the population size can be equally or even more rewarding. We approach the question in an empirical manner. We design and execute experiments for comparing the performance increase of a benchmark EA when augmented with self-adaptive control of parameters concerning selection and population size in isolation and in combination. With the necessary caveats regarding the test suite and the particular mechanisms used we observe that self-adapting selection yields the highest benefit (up to 30-40%) in terms of speed.

WILLEM: A Wireless InteLLigent Evacuation Method

In this paper we present WILLEM, a system for dynamic evacuation routing in buildings, using a wireless sensor network. Dynamic evacuation routing is the process of dynamically determining the fastest routes to the exits. The routes may be changed in case a fire occurs somewhere.We also present an algorithm for detecting congestions in corridors during evacuation, and a means of providing the people in those congestions an alternative route towards the exit. Each phase of the method is descibed extensively: the deployment of the wireless sensor network, the automatic topology learning of the network and the actual evacuation routing methods. We have built a simulation framework in which all types of evacuation routing can be simulated. The results of our experiments were surprising in the sense that dynamic evacuation routing turned out not to be faster than static evacuation routing in every setup; however, we did find out why this is the case. We also performed some experiments on a real wireless sensor network, in order to find out if our automatic configuration method could work in real life. The results are promising. We also present an algorithm for mapping the learned topology of the wireless sensor network upon a virtual map. This way, the network topology can be visualised – which is an important feature for emergency services.

Boosting Genetic Algorithms with Self-Adaptive Selection

In this paper we evaluate a new approach to selection in genetic algorithms (GAs). The basis of our approach is that the selection pressure is not a superimposed parameter defined by the user or some Boltzmann mechanism. Rather, it is an aggregated parameter that is determined collectively by the individuals in the population. We implement this idea in two different ways and experimentally evaluate the resulting genetic algorithms on a range of fitness landscapes. We observe that this new style of selection can lead to 30-40% performance increase in terms of speed.