Huiliang Zhang

A General Framework for Parallel BDI Agents

The traditional BDI agent has 3 basic computational components that generate beliefs, generate intentions and execute intentions. They run in a sequential and cyclic manner. This may introduce several problems. Among them, the inability to watch the environment continuously in dynamic environments may be disastrous. There is also no support for goal and intention reconsideration and consideration of relationships between goals at the architecture level. A parallel BDI agent architecture was proposed in [15] and evaluated in [16]. Based on the work in [15] and [16], we propose in this paper, a general framework for the parallel BDI agent model. Under this general framework, parallel BDI agents with different configurations depending on the availability of physical resources may be built. These agents have a number of advantages over the sequential one: 1. changes in the agents environment can be detected immediately; 2. emergencies will be dealt with immediately; 3. the support is provided at the architecture level for reconsideration of desires/intentions and the consideration of goal relationships when a new belief/desire is generated. We show some example parallel BDI agents with different configurations under the framework and their performance in a set of experiments.

Enabling Goal Oriented Action Planning with Goal Net

In this paper, we propose an agent planning system based on the Goal Net model. The agent’s goals are identified and organized in a composite goal hierarchy. Three kinds of relations between goals are defined: choice, concurrency and synchronization. Actions between goals are designed to accomplish subsequent goals. The agent’s desire is satisfied by accomplishing a serial of intermediary goals and finally achieving the ultimate goal that is satisfying the desire. The agent’s action plan is a list of actions to accomplish the intermediary goals in the solution. Because Goal Net is designed by considering agent’s possible desires directly, this bridges the distance between BDI agent design and the planning system. A searching algorithm is proposed to select goals in Goal Net. A case study of e-learning agent is shown to demonstrate the planning system.

Train Fuzzy Cognitive Maps by gradient residual algorithm

Fuzzy Cognitive Maps (FCM) is a popular technique for describing dynamic systems. A FCM for a dynamic system is a signed graph consisted of relevant concepts and causal relationships/weights between the concepts in the system. With suitable weights defined by experts in the related areas, the inference of the FCM can provide meaningful modeling of the system. Thus correctness of the weights is crucial to the success of a FCM system. Normally the weights are set by experts in the related areas. Considering the possible inefficiency and subjectivity of experts when judging the weights, it is an appealing idea to generate weights automatically according to the samples obtained through observation of the system. Some training algorithms were proposed. However, to our best knowledge, few learning algorithm has been reported to generate weight matrix based on sample sequences with continuous values. In this paper, we introduce a new learning algorithm to train the weights of FCM. In the proposed algorithm, the weights are updated by gradient descent on a squared Bellman residual, which is an accepted method in machine learning. The experiment results show that given sufficient training samples, the correct weights can be approximated by the algorithm. The algorithm proposes a new way for FCM research and applications.

How Fast Can a BDI Agent Respond

It is very important to know how fast a BDI agent can react to and process incoming event sequences if we want to apply such autonomous agents into time-sensitive applications like the Close-In weapon system in air-carriers. In [15], we proposed an analysis method for traditional sequential agents. In this paper we extend the theoretical analysis method to parallel BDI agents. Our method can estimate the average response time using the average attributes of a sequence of events based on probability and queueing theory. The simulation experiments show that our theoretical analysis method is effective. We also show by an experiment that an agent that dynamically allocates its computational time resources perform better than one that does not. Thus, the theoretical method suggests a way to quickly estimate the performance of an agent if the average attributes of the incoming event sequence are known in advance. Such an analysis of average response time can definitely benefit constructing more efficient BDI agents that situate in time-sensitive environments.

Motivated Learning for Goal Selection in Goal Nets

In Psychology, goal-setting theory, which has been studied by psychologists for over 35 years, reveals that goals play significant roles in incentive, action and performance for human beings. Based on this theory, the model of goal net has been proposed as a goal oriented agent model. The previous investigation has shown that the goal net model can support well multiple action and goal selection. In this paper, we will further show that the goal net model can simulate motivated learning of goal selections. More specifically, a reorganization algorithm is proposed to convert an original goal net to its counterpart that our learning algorithm can operate on. Our experiments show that in dynamic environments, agents with learning algorithms outperform agents with the recursive searching algorithm. In addition, the reorganization algorithm is not limited to the goal net model. It is applicable to other agent models.

An agent's activities are controlled by his priorities

Activity scheduling mechanism plays a critical role in the correct behaviour of BDI agents. For example, a robotic agent to serve at home should carry out the right activities at the right times. However the scheduling of deliberation about new beliefs and the scheduling of intention execution have not been carefully studied in most BDI systems. Usually if there is any differentiation of urgency among different tasks, a constant utility/priority value is used by a task selection fnction. We argue that priorities should be allowed to change with time and a linear function of time may not be the best for all tasks. In this paper, we propose to enrich the BDI framework with an extension which consists of 2 processing components, a PCF (Priority Changing Function) Selector and a Priority Controller. With this extension priorities of desires/intentions may have different initial values and may be changed with time according to the chosen PCFs. We propose a method of constructing PCFs which model the change of priorities in human behaviors when dealing with several things at the same time. We also propose a method to realize the change of the priorities of existing desires/intentions due to the generation of new beliefs/desires/intentions if necessary. We show by simulation experiments that Ramp function and especially the Sigmoid function can control the activities of an agent better than constant priorities with respect to getting tasks of various importance and urgency done with smaller Mean Earliness and smaller Mean Tardiness.

Activity scheduling for a robotic caretaker agent for the elderly

A real-time robotic agent that takes care of an elderly person at home will need to schedule various tasks in real time. The deadlines of its tasks are generally soft (missing a deadline by a few minutes in most cases has no serious consequences). Another characteristic is that many tasks are preferably done close to some time points instead of as soon as possible. To support such time management behaviour, we propose to enrich the BDI agent framework with an extension which consists of two processing components, a priority changing function (PCF) selector and a priority controller. The priorities of desires/intentions are represented by their PCFs. A PCF is a function of both time and the utility value of a desire/intention. So it represents both the urgency and the importance (beneficial value) of a desire/intention. We propose a method of constructing PCFs which model the change of priorities of tasks as time passes. Simulation experiments show that sigmoid function can control the activities of an agent better than constant priorities with respect to getting tasks done with smaller mean earliness and smaller mean tardiness. A BDI agent built with this time management mechanism will try to complete its tasks at the right time. The order in which multiple goals and multiple intentions are handled will be flexible and time dependent.