Enhancing Inference in Relational Reinforcement Learning via Truth Maintenance Systems

Size: px
Start display at page:

Download "Enhancing Inference in Relational Reinforcement Learning via Truth Maintenance Systems"

Transcription

1 Enhancing Inference in Relational Reinforcement Learning via Truth Maintenance Systems Mandana hamidi Tele Robotics and Applications Department Italian Institute of Technology (IIT) Genova, Italy Amir Fijany Tele Robotics and Applications Department Italian Institute of Technology (IIT) Genova, Italy Jean-Guy Fontaine Tele Robotics and Applications Department Italian Institute of Technology (IIT) Genova, Italy Abstract Computational complexity is still a challenging problem for intelligent systems operating in complex environments. To tackle it, an agent has to deal with perceptual information intelligently. In this paper, we propose an efficient and adaptive reasoning system based on Adaptive Logic Interpreter reasoning system, a mechanism for guiding inference through relational reinforcement learning, and a variation of truth maintenance systems to speed up the inference. Relational reinforcement learning guides the inference toward the most rewarding parts of the knowledge base and truth maintenance system maintains beliefs, avoids repetitive inferences and reduces the state space. Empirical results demonstrate higher performance than the basic approach in terms of number of inferred instances, average reward, and average reward accuracy. Keywords-component; Inference engine; truth maintenance system; relational reinforcement learning; Dynamic Environment. I. INTRODUCTION Logical reasoning systems have been widely used for solving problems from different applications domains. However, these systems are usually limited by the computational complexity of their inference process, particularly when the number of inferential rules and amount of available background knowledge grow. The inference mechanism is the core element of an artificial intelligent agent. When such an agent operates in real-time, it is subject to time and computational resource constraints. An agent like a robot living in a complex perceptual environment must guarantee a reasonable response time given its computational resources. One common approach to overcome the complexity bottleneck is to guide the reasoning process toward the most rewarding parts of the knowledge base which leads to a better performance with time and computational constraints [1]. It is clear that under time and computational constraints an agent cannot try all possible inferences. Therefore a viable approach is to give priority to drawing more important conclusions first and then, if time permits, move on to other less important ones. Given a criterion, the question now is how to focus the reasoning of a reactive agent which cannot afford all possible inferences and therefore have to assign its time and computational resources efficiently? In another words how to attend to the most important, in terms of reward, parts of the knowledge base? Some proposed methods attempted to speed up the inference mechanism by combining some clever indexing schemes with exhaustive search methods. For instance, Rete match algorithm, operate by constructing a network to determine which rules should be triggered [3]. Truth Maintenance Systems (TMSs) applied a collection of techniques for maintaining beliefs and doing belief revision [2]. In our work, we combine a variation of TMSs with learning algorithm to gain more efficient inference system. Some approaches have incorporated heuristic and control strategies to guide inference. The meta-level reasoning system developed by Genesereth et al. [4] lets the designer write Prolog-like control clauses that specify how inference rules should be prioritized. Other works, have attempted to apply metareasoning, reasoning about reasoning, to deal with real-world complexity. For instance, Russell and Welford [5] sought to develop a general framework for meta-reasoning based on probability and decision theory. They introduced the idea of rational metareasoning, wherein computations are mental actions. Other approaches have focused on learning control rules to reduce search or speed up processing [6]. In [7] explanation-based learning techniques were combined with inductive logic programming ideas to learn such control conditions over inference rules. In a different approach, Cohen et al. [8] used bootstrapping to learn similar rules. These approaches assumed a query-based, top-down inference mechanism and sought to modify the logic program itself in order to achieve performance gain. In [9], authors demonstrated that it is possible to learn to guide inference efficiently via reinforcement learning. One of the learning methods which has received a lot of attentions since early 9 s is Relational Reinforcement learning (RRL), which combines the reinforcement learning with relational learning or inductive logic programming [1]. In the RRL algorithm that we applied in this work, similar to Russell and Welford, actions are internal to the agent. In recent years, there has been an increased interest in applying learning methods for reasoning systems under

2 resource constraints. An efficient approach for speeding up inference is Adaptive Logic Interpreter () reasoning system [1] that employs a variation of RRL for controlling the inference process. In this system, the reasoning is focused on the most rewarding parts of its knowledge base and hence performs better under time and computational resource constraints [1]. This approach assumes a disembodied agent interacting with its surrounding world. However, when the time constraint is extremely strict or the environment is changing too rapidly, does not provide a satisfying performance gain. This is partly because RRL s training is time consuming, especially when the state space is large, and partly because of the fact that it does not maintain its beliefs and makes inference regardless of whether it has been already inferred in the previous steps or not. To tackle the above-mentioned drawbacks, in this paper Justification-based TMS (JTMS) is applied for enhancing and making it efficient in real time situations. We called our approach, which combines the desirable features of both JTMS and and diminishes their disadvantages. Our approach differs from in two primary aspects. Firstly, we employ a JTMS, to maintain the beliefs and their dependencies. Secondly, JTMS helps RRL to reduce (prune) the space of actions to consider during learning. To the best of our knowledge, this is the first learning method combining RRL and JTMS. This paper is organized as follows. In Sections 2 and 3, and TMSs approaches are briefly discussed. Our new approach for combining JTMS and is presented in Section 4. Experiments and performance results are discussed in Section 5. Finally, some concluding remarks are made in Section 6. II. ADAPTIVE LOGIC INTRPRETER () REASONING SYSTEM is an online learning algorithm which enables an agent architecture acquire a controlled inference strategy adapted to the environment [1]. It is built upon ICARUS, a reactive agent architecture which supports reasoning and decision making [13]. ICARUS controls a physical agent in the environment composed of collections of objects whose attributes and mutual relations change over time. The agent s knowledge about domain content in terms of concepts and skills are stored in the long-term memory. Concepts describe situations in the environment and skills describe how to respond to these situations. The agent s dynamic beliefs are stored in short-term conceptual memory. These beliefs are specific instances of concept definitions that can be inferred from the perceptual buffer. Table I shows three concept definitions from blocks-world environment with variables indicated by question mark. Each concept can have five optional fields :percept(the perceived entities), :positives(the lower-level concepts that must match), :negatives(the lowerlevel concepts that must not match), :tests (the numeric relations that must hold), and :reward (the internal reward function for the matched concept). ICARUS operates in a cyclic fashion. In each cycle, the perception of the agent changes depending on its field of view. Then its perceptual memory is updated with the descriptions of all objects that are visible in the environment. An inference mechanism revises the agent s belief based on the perceptions and its domain knowledge. ICARUS infers all the matched instances of concepts in the hierarchy in a breath-first, bottom-up fashion. Finally, based on the belief state, the agent finds all the applicable skills and selects the skill with highest utility to execute. TABLE I. SOME CENCEPT DEFINITIONS FORM BLOCKS-WOLRD. (is-block (?b) :percepts ( (block?b xpos?x) ) :reward.) (left-of (?b1,?b2) :percepts ((block?b1 xpos?x1) (block?b2 xpos?x2)) :positives ( (is-block?b1) (is-block?b2)) :tests ( (<?x1?x2)) :reward.) (between (?b1,?b2,?b3) :percepts( (block?b1 xpos?x1) (block?b2 xpos?x2) (block?b3 xpos?x3)) :positives ( (left-of?b1?b2) (left-of?b2?b3) ) :test ((<?x3-x1 3)) :reward (*1(-3?x3?x2?x1) ) ICARUS s inference method, exhaustive inference, considers concepts in bottom-up breath-first fashion with no control over the reasoning strategy. Hence, it does not perform well when the agent operates under time constraints. In order to overcome the drawback of exhaustive inference, applies RRL to capture the important instances by estimating their values. In all the concepts are instantiated but only the most important ones are inferred. consists of two components: a) learning mechanism and b) generalization. The first component uses attentionrelevant values assigned to instances to determine the most rewarding subset. The second component generalizes the instance-specific values that result from the reinforcement learning algorithm to value functions for their corresponding concept definitions. A. Learning Mechanism Let be the set of all concept instances. The agent aims to find the subset which maximizes the accumulative reward under time constraint. To achieve this goal, the RRL algorithm tries to learn values over. Mental state is defined as the set of instances inferred to be true after inference steps within the current execution cycle. Fringe is defined as the set of all inferable concept instances whose children are already included in state. Each inference step consists of selecting an inferable instance from and checking whether it holds or not. A, is an inference step that infers the instance, at the current step. After inferring, the reward of the action is computed by (1): Where is the reward function, associated with concept instance and denotes the attribute vector for perceived

3 objects on which depends. The expected value of instance is updated by (2): Where is a set of inferable instances in the current fringe for which is a child. Furthermore, and are defined as: Where visits( ) indicates the number of updates performed on. For better understanding, suppose the knowledge base consists of three concept definitions introduced in Table I. Fig. 1 (bottom) shows how the belief state and the fringe are updated by selecting a valid action. Once the system selects instance (is-block C) as an action and infers it as a true belief, and adds it to the belief state. Because (is-block B) and (is-block A) are already in the belief state, the new inferable instances (left-of A C), (left-of C A), (left-of B C), and (left-of C B) are added to the fringe. (left-of A B) (left-of B A) A B C Infer (is-block C) (left-of AB) (left-of BA) (left-of AC) (left-of CA) (left-of BC) (left-of CB) Figure 1. Top) A simple example with three blocks from blocks-world. Bottom) Belief state and fringe after updating. The bold literals are the instances belonging to belief states, and the blue literals are the instances belong to fring. B. Generalization mechanism For each concept definition, generalization algorithm uses linear regression method to initialize the expected value of new candidate beliefs in the instance space. The linear regression method,, is incrementally updated by (4) and is fitted to the training examples. Here x(u) denotes the vector of attributes of the perceptions that appear in concept instance u, and is the set of all instances derived from concept definition c. III. TRUTH MAINTENANCE SYSTEMS(TMSS) In 1987, Doyle introduced TMS, a general problem solving facility to help inference engine efficiently manipulate beliefs [2]. In most problem solvers, inference engine explores alternatives, makes belief and examines the consequence of beliefs. But it often makes repetitive inferences. TMS maintains the caches of all the inferences ever made. Thus inferences, once made, need not be repeated and contradictions, once discovered, are avoided in the future. During problem solving, the inference engine and TMS continuously interact through a well-defined protocol, as shown in Fig. 2. Whenever the problem solver performs an inference step, it sends a datum to the TMS. The task of TMS is to maintain a consistent state of belief in such a way that they are not known to be contradictory and no belief is kept without reason. Inference engine Figure 2. Structure of truth minatenance systems TMSs can be classified as JTMS [2], Assumption-based TMS [11], and Logic-based TMS. These systems differ in the way they represent dependencies between the derived facts and the facts they have been derived from [12]. JTMS, the simplest kind of TMSs, is classically represented by a network of nodes together with a set of links (justifications) that represent dependencies between nodes. Every important problem solving datum (fact, assumption, and justification) sent to the TMS, is stored as a node in TMS. Assumptions nodes represent the data which the inference engine prefers to believe but which it may want to retract later. Each justification records the dependencies between the justified nodes and their antecedents. The current belief status of a node is represented as a label. If nodes are currently believed their labels are in and if they are currently not believed their labels are out. A justification contains two lists - inlist and outlists - to describe which nodes must be in and which must be out for a justified node to be believed. Justifications are used for two reasons. The first is to perform the belief update when the belief state of a node changes. Justifications are used to find all the affected nodes, whose belief states are then reexamined. As a result, some of these nodes will become in, while, others will become out. Second, justifications are used to handle contradiction. When a contradiction is discovered by a problem solver, a justification for the contradiction is added. This system searches through the dependency network (using justifications and nodes) for the assumptions underlying the contradiction. It selects one assumption as the culprit and justifies one of the nodes from its outlist, thus removing the contradiction. IV. Justifications Assumptions Beliefs Contradictions TMS TMS-BASED Our approach is similar to in terms of learning, since both methods learn by RRL; the difference is that we

4 apply JTMS for improving the performance. JTMS is employed not only for maintaining the beliefs, but also for reducing RRL s state space. The schematic of is exhibited in Fig. 3. It consists of inference engine, long-term conceptual memory, JTMS, learner and perception. Despite s model that stores agent s dynamic beliefs in short-term conceptual memory, in our model beliefs and their dependencies are stored as JTMS nodes. While in the set of all concept instantiations are kept in memory and are updated whenever a new perception is added or an old one is deleted, in our model, concept instances are incrementally generated and added to JTMS. Long-term Conceptual Memory Perception Environment Inference Engine Justifications/ Assumptions Inferable Instance Reward Beliefs/ Contradictions JTMS Figure 3. A schematic of Learner (RRL) The set of all nodes or concept instances,, in JTMS is classified in three set of nodes:, and. is the set of believed JTMS nodes with label, is the set of unbelieved JTMS nodes with label, and is the set of nodes in JTMS that have not inferred and the belief status of these nodes are unknown. works in a cyclic fashion. Each cycle consists of two steps: A) Perception and Belief revision, and B) learning mechanism. A. Perception and belif revision At the beginning of each cycle, the agent perceives the environment and updates its perceptual memory. Then a matcher, checks those primitive concepts that are matched with the perceptual buffer information. It generates primitive concept instances and sends them to inference engine. Inference engine checks these new beliefs with JTMS. If the new belief has not been seen before by JTMS, then stores it as an assumption node in JTMS. All the dependencies between new belief and other beliefs are sent to the JTMS and added to. If the system perceives an object as deleted, it will send a justification for this contradiction to the JTMS, and JTMS automatically revises its beliefs by searching through the dependency network (using justifications and nodes) for the assumptions underlying the contradiction and then changing their labels form in to out. B. Learning Mechanism This step is similar to s learning algorithm. The only difference is the way that JTMS helps RRL in updating fringe. Whenever RRL selects an inferable instance u from fringe and sends it to inference engine, inference engine asks JTMS, whether this data has been already seen or not. If u is a new inferable concept instance it will be then inferred by the inference engine and will be stored in JTMS. If it has been inferred before, it does not need to be inferred again. Once instance u is inferred true, all JTMS nodes with label in or unknown whose child is u, are added to the fringe. More precisely, the fringe does not contain any nodes with label out. Therefore, it reduces the size of the fringe and prevents the selection of concept instances that have been already inferred false on previous cycles. TMS-based algorithm is summarized in Table II. TABLE II. ALGORITHM FOR TMS-BASED Repeat the following steps for N execution cycles 1. Perception and belief revision 1.1. Let state and If a new object is perceived, A new assumption node with label is stored in JTMS ( ) All inferable instances whose children are included in n and other assumption node, are saved in with value zero If a perceived object is deleted, add a justification for this contradiction to JTMS Initialize the fringe with all JTMS assumption nodes. 2. Learning mechanism: Repeat the following steps until the time runs out or there is no inferable instance in 2.1. Let ; select instance 2.2. Let and = 2.3. If, add it to,and return reward 2.4. If, infer u If is true, add it to, compute its reward, change label to. Add all inferable instances whose children are included in nodes of, and are not exist in JTMS, to If is false, set its reward to zero and justification for this contradiction to JTMS Remove from Let 2.7. Update by using (2),(3) and For a better understanding, let us assume the simple blocks-world example shown in Fig.1, top). In the first cycles, there is no any concept instance in JTMS. Therefore, RRL generates all possible inferable instances for updating its fringe. A sample of this updating was shown in Fig. 1(bottom). After few cycles, inference engine infers that concept instances (left-of A, B), (left-of B, C), and (left-of A, C) are false. Thus, these inference concepts are stored in JTMS with label out. RRL can apply this declarative knowledge for updating the fringe. Comparison of Fig.1 (bottom) and Fig. 4, shows that number of member of the fringe is reduced from 6 to 3. In such a case, RRL instead of

5 Number of inferred instances checking 6 inferable instances in fringe, checks only 3 instances. (left-of A B) (left-of AB) (left-of AC) (left-of BC) Infer (is-block C) Figure 4. An example of belief state and fringe update in TMS-based. The bold literals are the instances belong to belief states, and the blue literals are the instances belong to fringes V. EXPERIMENT AND RESULTS In order to evaluate the performance of the proposed algorithm, we have carried out a set of experiments employing a known blocks-world environment [1]. This environment contains a set of blocks placed in line, each with a name and a position specified as its distance from a reference point, which we call the origin. Three concepts for this environment are shown in Table I. Clearly, when a large number of blocks are presented, the number of feasible instances from between concept will be enormous. It is assumed that the agent is located at the origin and has a tendency to interact with nearby blocks. Therefore, we assigned a reward function to the highest level concept (between) as a linear function that favors relations whose corresponding blocks are closer to the origin. The other two concept definitions (left-of and is-block) have no assigned reward function. There are several factors that affect the performance of inference systems, e. g., the rate of the change in the environment, domain complexity, and time limit for each inference cycle. In the following, we discuss four experiments to investigate the effect of these factors on the performance of five different inference systems: TMS-based, with generalization,, with generalization, and exhaustive inference. Reported results are averaged over 2 independent runs. A. Experiment I First, we analyze the number of inferred instances on five different inference systems. The number of inferred instances refers to the number of concept instances, which are inferred by the inference engine in one cycle. Six blocks A, B, C, D, E, and F with positions 1, 2, 3, 4, 5, and 6 are placed in initial world. Each inference system learns for 1 cycles under a fixed time limit of.4 seconds. Then they are tested by adding and deleting blocks from the environment. Eight blocks G, H, I, J, K, L, M, and N with positions.5, 6, 1.7, 7, 1.5,.25, 8, 1, and 2.5 are added to the environment at cycles 1, 125, 125, 15, 175, 2, 225, and 275 respectively. Blocks A and D are deleted from the environment at cycles 16 and 25, respectively. Fig. 5 summarizes the number of inferred instances over cycles. As expected, in both and TMSbased with generalization, when a new object is added to the environment, the number of inferred instances jump for a few cycles. The reason is that, in these cycles, the inference engine infers new instances, which have not yet been inferred. When JTMS caches these instances, the number of inferred instances decreases again. In the other inference systems, the number of inferred instances in each cycle is higher than the two above mentioned methods. This result clearly demonstrates that the JTMS plays an important role in decreasing the number of inferences in the system. In, with generalization, and exhaustive inference systems, as the number of objects increases, the number of inferred instances decreases. When an object is deleted from environment (in cycles 16 and 25), the number of inferences is increased. It shows that, the number of inferred instances in each cycle, is related to the complexity of the environment. Notice, a good inference system not only should avoid repetitive inferences, but also should be able to infer new belief instances, when new objects are added to the environment. Generalization mechanism, by initializing the expected value of new belief candidates, plays an important role in dealing with new inferred instances. It causes these beliefs to have more chance to be selected as an action. By comparing number of inferred instances in TMS-based and with generalization, one can understand that at cycles 2, 225, and 275, when objects L, M, and N are added to the environment, TMSbased with generalization is more sensitive to the change of the environment. Although, has less number of inferred instances w ith genralization w ith genralization Number of cycles Figure 5. Comparing the number of infered instancesof five inference systems over cycles. To see how successful our inference system is in guiding inference toward the most rewarding parts of the current instance space, we measure the reward accuracy. Reward accuracy is defined as and is measured in every cycle. denotes the cumulative rewards obtained by the inference system under time constraints in each cycle. denotes the total reward that would be accumulated in the same cycle by making all possible inferences under no time constraint. The reward accuracy of five inference systems over cycles is compared in Fig. 6. From cycle one to cycle 1, all inference systems achieved maximum accuracy

6 Average reward Reward Reward accuracy Average reward accuracy reward. It shows that the inference systems have enough time to generate all possible concept instances in each cycle. However, after cycle 1, by increasing the number of objects in the environment, the size of the belief state is increased and there is not enough time to generate all inference instances. Since exhaustive inference system works in the bottom-up breath-first manner, it has no time to infer higher-level concept instances. Therefore, the reward accuracy remains zero after cycle 1. The only method that its reward accuracy does not decrease rapidly is the TMSbased with generalization. 1 cycles. The averaged inferred reward during the last 1 cycles over decreasing amounts of variable time, is plotted in Fig. 8. Average reward accuracy drops as time limit increases (smaller values). with generalization gains the best reward accuracy among the other methods. with generalization gives better performance than and exhaustive inference system w ith genralization w ith genralization Number of cycles Figure 6. Accuracy of diferrent inferences systems over cycles Reward of the inference systems over cycles is illustrated in Fig. 7. Here, the reward refers to the cumulative reward obtained on a particular execution cycle. Exhaustive inference performs very poorly,, with generalization, and give better performance and with generalization performs the best. This is due to better action selection and value initialization of with generalization, which eventually ends up to higher reward w ith genralization w ith genralization Total rew ard in the environemnt Time limitation for each inference cycle(seconds) Figure 8. Reward accuracy of different inference systems over increasing amounts of available time for each inference cycle. C. Experiment III We have also considered the effect of domain complexity, which refers to the size of the instance space. Similar to the experiment II, five inference systems are trained in the similar situation. But, at cycle 2, the complexity is increased by abruptly adding p blocks, with p a random number between one to eight. As before, we measured the average inferred reward over the final1 cycles. Fig 9 shows the average reward results as a function of the number of blocks added. As expected, when the environment becomes complex, and with generalization perform better than the other methods. ALDIN and with generalization perform poorly. system does not achieve any reward with genralization with genralization with genralization with genralization Number of cycles Figure 7. reward comparison over cycles. B. Experiment II Next, the effect of time constraint on the performance of the five different inference systems is considered. Each system learns for 2 cycles under a fixed time limit of.4 seconds in an initial world state with six blocks. Then the systems are tested under various time limits ranging from.1 to.2 seconds, in a dynamic environment. Every 25 cycles, a new block is added to the environment for a total of Number of blocks added at cycle 2 Figure 9. Performance comparision between five inference sytems over increasingly complex blocks-world environments. D. Experiment IV Finally, we have studied the rate of the change in the environment. The train phase of this experiment is similar to

7 Average Reward Accuracy experiment II. After a 2-cycle learning period, we inserted new blocks one at a time with a specific number of cycles between consequent insertions. Then, the average performance of each inference system over its last 1 cycles is measured. Fig. 1 shows and TMSbased with generalization achieves the best performance among other methods with genralization with genralization Number of cycles betw een consequent block insertions Figure 1. Accuracy of diferrent inference systems over increasind rates of environemntal change in the block-world domain When the environment changes rapidly, when it becomes too complex, or when the available time for each inference cycle is too limited, the performance of all inference mechanisms degrade significantly. Indeed, they don t have enough time to learn the value of all possible concept instances. However, even in the short learning time, TMSbased systems gain better performance than systems for two main reasons. First, does not cache the results of the inferences. Second, in, the concept instances that have been inferred false in the previous cycles, might be selected in current cycle. Selecting these instances does not have any positive effect on the performance of the systems and it only consumes the time of the learning. VI. CONCLUSION In this paper, we presented, a new approach for adaptive reasoning systems based on system. We applied JTMS to improve the performance of. The advantages of JTMS are two-fold: first, maintaining the beliefs to reduce the number of inferred instances and to speed up inference mechanism; and second, reducing the size of the state space that needs to be considered by the RRL method. Our experiments on dynamic blocks-world domain proved that TMS-based achieved the best results in guiding the reasoning process towards high-utility parts. Our results also demonstrate that learns much faster than and exhaustive inference systems in dynamic environment. This clearly proves that JTMS is indeed a good candidate for speeding up the s inference. For future work, we plan to investigate the role of the TMS in guiding the leaning algorithm. TMSs can have more collaboration with learning mechanism. Whenever the environment changes, TMS updates the truth-value of beliefs. It can guide RRL to concentrate on the belief instances that their truth-values have been updated recently by TMS and need to be inferred again. Since uncertainty is an inevitable fact for real-world applications, we intend to extend our approach to probabilistic domain. One promising direction in this regard is applying it in to more complex task like visual attention and human robot interaction. Furthermore, JTMS is restricted in a sense that it can only accept Horn clauses as justifications. However, many applications need to express more than just Horn clauses, and hence we plan to investigate other type of TMS that which accept general clauses. VII. AKNOWLEDGEMNT The authors would like to thank Nima Asgharbeygi for helpful discussion on algorithm. REFERENCES [1] N. Asgharbeygi, N. Nejati, P. Langley, and S. Arai, Guiding inference through relational reinforcement learning, In: Inductive Logic Programming: 15th International Conference, Springer, Germany, 25, pp [2] J. Doyle, A truth maintenance system, Artificial Intelligence. vol. 12, pp ,1979. [3] C. Forgy, Rete: A fast algorithm for the many pattern/many object pattern match problem, Artificial Intelligence 19, pp , [4] M. R. Genesereth and M.L. Ginsberg, Logic programming. Communications of the ACM, vol. 28, PP ,1985. [5] S. Russell and E. Wefald, Principles of metareasoning, In: Proceedings of the First In- ternational Conference on Principles of Knowledge Representation and Reasoning, San Mateo, CA, Morgan Kaufmann,1989. [6] S. Minton, Quantitative results concerning the utility of explanationbased learning, Artificial Intelligence, vol. 42, pp ,199. [7] J.M. Zelle and R.J. Mooney, Combining FOIL and EBG to speedup logic programs, In: Proceedings of the Thirteenth International Joint Conference on Artificial Intelligence, Chambery, France, Morgan Kaufmann, pp , [8] W.W. Cohen and Y. Singer, A simple, fast, and effective rule learner, In: Proceedings of the Forteenth National Conference on Artificial Intelligence. pp , [9] M.E. Taylor, C. Matuszek, R. Smith,and M. J. Witbrock, Guiding Inference with Policy Search Reinforcement Learning, FLAIRS Conference, pp , 27. [1] S. Dzeroski, L. De Raedt, and H. Blockeel,. Relational reinforcement learning; In Proceedings of the Fifteenth International Conference on Machine Learning, pp , Madison, WI, Morgan Kaufmann [11] J. De Kleer, An assumption based TMS, Artificial Intelligence,.vol. 28, pp , [12] M. Stanojevic, S. Vranes, and D. Velasevic, Using Truth Maintenance Systems: A Tutorial, IEEE Expert: Intelligent Systems and Their Applications. vol. 9, pp , [13] D. Choi, M. Kaufman, P. Langley, N. Nejati, and D. Shapiro, An architecture for persistent reactive behavior, In Proceedings of the Third International Joint Conference on Autonomous Agents and Multi Agent Systems, ACM Press, pp ,New York, 24.

Guiding Inference through Relational Reinforcement Learning

Guiding Inference through Relational Reinforcement Learning Guiding Inference through Relational Reinforcement Learning Nima Asgharbeygi, Negin Nejati, Pat Langley, and Sachiyo Arai Computational Learning Laboratory Center for the Study of Language and Information

More information

Plan Recognition through Goal Graph Analysis

Plan Recognition through Goal Graph Analysis Plan Recognition through Goal Graph Analysis Jun Hong 1 Abstract. We present a novel approach to plan recognition based on a two-stage paradigm of graph construction and analysis. First, a graph structure

More information

Handling Partial Preferences in the Belief AHP Method: Application to Life Cycle Assessment

Handling Partial Preferences in the Belief AHP Method: Application to Life Cycle Assessment Handling Partial Preferences in the Belief AHP Method: Application to Life Cycle Assessment Amel Ennaceur 1, Zied Elouedi 1, and Eric Lefevre 2 1 University of Tunis, Institut Supérieur de Gestion de Tunis,

More information

Grounding Ontologies in the External World

Grounding Ontologies in the External World Grounding Ontologies in the External World Antonio CHELLA University of Palermo and ICAR-CNR, Palermo antonio.chella@unipa.it Abstract. The paper discusses a case study of grounding an ontology in the

More information

Plan Recognition through Goal Graph Analysis

Plan Recognition through Goal Graph Analysis Plan Recognition through Goal Graph Analysis Jun Hong 1 Abstract. We present a novel approach to plan recognition based on a two-stage paradigm of graph construction and analysis. First, a graph structure

More information

A Computational Account of Everyday Abductive Inference

A Computational Account of Everyday Abductive Inference A Computational Account of Everyday Abductive Inference Will Bridewell Center for Biomedical Informatics Research Stanford University, Stanford, CA 94305 Pat Langley Computing Science and Engineering Arizona

More information

LECTURE 5: REACTIVE AND HYBRID ARCHITECTURES

LECTURE 5: REACTIVE AND HYBRID ARCHITECTURES Reactive Architectures LECTURE 5: REACTIVE AND HYBRID ARCHITECTURES An Introduction to MultiAgent Systems http://www.csc.liv.ac.uk/~mjw/pubs/imas There are many unsolved (some would say insoluble) problems

More information

Coherence Theory of Truth as Base for Knowledge Based Systems

Coherence Theory of Truth as Base for Knowledge Based Systems Association for Information Systems AIS Electronic Library (AISeL) AMCIS 1996 Proceedings Americas Conference on Information Systems (AMCIS) 8-16-1996 Coherence Theory of Truth as Base for Knowledge Based

More information

Programming with Goals (3)

Programming with Goals (3) Programming with Goals (3) M. Birna van Riemsdijk, TU Delft, The Netherlands GOAL slides adapted from MAS course slides by Hindriks 4/24/11 Delft University of Technology Challenge the future Outline GOAL:

More information

Towards Learning to Ignore Irrelevant State Variables

Towards Learning to Ignore Irrelevant State Variables Towards Learning to Ignore Irrelevant State Variables Nicholas K. Jong and Peter Stone Department of Computer Sciences University of Texas at Austin Austin, Texas 78712 {nkj,pstone}@cs.utexas.edu Abstract

More information

Chapter 2. Knowledge Representation: Reasoning, Issues, and Acquisition. Teaching Notes

Chapter 2. Knowledge Representation: Reasoning, Issues, and Acquisition. Teaching Notes Chapter 2 Knowledge Representation: Reasoning, Issues, and Acquisition Teaching Notes This chapter explains how knowledge is represented in artificial intelligence. The topic may be launched by introducing

More information

A Comparison of Collaborative Filtering Methods for Medication Reconciliation

A Comparison of Collaborative Filtering Methods for Medication Reconciliation A Comparison of Collaborative Filtering Methods for Medication Reconciliation Huanian Zheng, Rema Padman, Daniel B. Neill The H. John Heinz III College, Carnegie Mellon University, Pittsburgh, PA, 15213,

More information

Positive and Unlabeled Relational Classification through Label Frequency Estimation

Positive and Unlabeled Relational Classification through Label Frequency Estimation Positive and Unlabeled Relational Classification through Label Frequency Estimation Jessa Bekker and Jesse Davis Computer Science Department, KU Leuven, Belgium firstname.lastname@cs.kuleuven.be Abstract.

More information

The Semantics of Intention Maintenance for Rational Agents

The Semantics of Intention Maintenance for Rational Agents The Semantics of Intention Maintenance for Rational Agents Michael P. Georgeffand Anand S. Rao Australian Artificial Intelligence Institute Level 6, 171 La Trobe Street, Melbourne Victoria 3000, Australia

More information

Perception Lie Paradox: Mathematically Proved Uncertainty about Humans Perception Similarity

Perception Lie Paradox: Mathematically Proved Uncertainty about Humans Perception Similarity Perception Lie Paradox: Mathematically Proved Uncertainty about Humans Perception Similarity Ahmed M. Mahran Computer and Systems Engineering Department, Faculty of Engineering, Alexandria University,

More information

Identifying Parkinson s Patients: A Functional Gradient Boosting Approach

Identifying Parkinson s Patients: A Functional Gradient Boosting Approach Identifying Parkinson s Patients: A Functional Gradient Boosting Approach Devendra Singh Dhami 1, Ameet Soni 2, David Page 3, and Sriraam Natarajan 1 1 Indiana University Bloomington 2 Swarthmore College

More information

Using Bayesian Networks to Direct Stochastic Search in Inductive Logic Programming

Using Bayesian Networks to Direct Stochastic Search in Inductive Logic Programming Appears in Proceedings of the 17th International Conference on Inductive Logic Programming (ILP). Corvallis, Oregon, USA. June, 2007. Using Bayesian Networks to Direct Stochastic Search in Inductive Logic

More information

ICS 606. Intelligent Autonomous Agents 1. Intelligent Autonomous Agents ICS 606 / EE 606 Fall Reactive Architectures

ICS 606. Intelligent Autonomous Agents 1. Intelligent Autonomous Agents ICS 606 / EE 606 Fall Reactive Architectures Intelligent Autonomous Agents ICS 606 / EE 606 Fall 2011 Nancy E. Reed nreed@hawaii.edu 1 Lecture #5 Reactive and Hybrid Agents Reactive Architectures Brooks and behaviors The subsumption architecture

More information

Lecture 2: Foundations of Concept Learning

Lecture 2: Foundations of Concept Learning Lecture 2: Foundations of Concept Learning Cognitive Systems - Machine Learning Part I: Basic Approaches to Concept Learning Version Space, Candidate Elimination, Inductive Bias last change October 18,

More information

Improving the Accuracy of Neuro-Symbolic Rules with Case-Based Reasoning

Improving the Accuracy of Neuro-Symbolic Rules with Case-Based Reasoning Improving the Accuracy of Neuro-Symbolic Rules with Case-Based Reasoning Jim Prentzas 1, Ioannis Hatzilygeroudis 2 and Othon Michail 2 Abstract. In this paper, we present an improved approach integrating

More information

ERA: Architectures for Inference

ERA: Architectures for Inference ERA: Architectures for Inference Dan Hammerstrom Electrical And Computer Engineering 7/28/09 1 Intelligent Computing In spite of the transistor bounty of Moore s law, there is a large class of problems

More information

1 What is an Agent? CHAPTER 2: INTELLIGENT AGENTS

1 What is an Agent? CHAPTER 2: INTELLIGENT AGENTS 1 What is an Agent? CHAPTER 2: INTELLIGENT AGENTS http://www.csc.liv.ac.uk/ mjw/pubs/imas/ The main point about agents is they are autonomous: capable of acting independently, exhibiting control over their

More information

Citation for published version (APA): Geus, A. F. D., & Rotterdam, E. P. (1992). Decision support in aneastehesia s.n.

Citation for published version (APA): Geus, A. F. D., & Rotterdam, E. P. (1992). Decision support in aneastehesia s.n. University of Groningen Decision support in aneastehesia Geus, Arian Fred de; Rotterdam, Ernest Peter IMPORTANT NOTE: You are advised to consult the publisher's version (publisher's PDF) if you wish to

More information

Positive and Unlabeled Relational Classification through Label Frequency Estimation

Positive and Unlabeled Relational Classification through Label Frequency Estimation Positive and Unlabeled Relational Classification through Label Frequency Estimation Jessa Bekker and Jesse Davis Computer Science Department, KU Leuven, Belgium firstname.lastname@cs.kuleuven.be Abstract.

More information

Fuzzy Decision Tree FID

Fuzzy Decision Tree FID Fuzzy Decision Tree FID Cezary Z. Janikow Krzysztof Kawa Math & Computer Science Department Math & Computer Science Department University of Missouri St. Louis University of Missouri St. Louis St. Louis,

More information

On the Representation of Nonmonotonic Relations in the Theory of Evidence

On the Representation of Nonmonotonic Relations in the Theory of Evidence On the Representation of Nonmonotonic Relations in the Theory of Evidence Ronald R. Yager Machine Intelligence Institute Iona College New Rochelle, NY 10801 Abstract A Dempster-Shafer belief structure

More information

A Framework for Medical Diagnosis using Hybrid Reasoning

A Framework for Medical Diagnosis using Hybrid Reasoning A Framework for Medical using Hybrid Reasoning Deepti Anne John, Rose Rani John Abstract The traditional method of reasoning was rule-based reasoning (). It does not use past experiences to reason. Case-based

More information

Support system for breast cancer treatment

Support system for breast cancer treatment Support system for breast cancer treatment SNEZANA ADZEMOVIC Civil Hospital of Cacak, Cara Lazara bb, 32000 Cacak, SERBIA Abstract:-The aim of this paper is to seek out optimal relation between diagnostic

More information

Paper presentation: Preliminary Guidelines for Empirical Research in Software Engineering (Kitchenham et al. 2002)

Paper presentation: Preliminary Guidelines for Empirical Research in Software Engineering (Kitchenham et al. 2002) Paper presentation: Preliminary Guidelines for Empirical Research in Software Engineering (Kitchenham et al. 2002) Who? Eduardo Moreira Fernandes From? Federal University of Minas Gerais Department of

More information

Intelligent Agents. CmpE 540 Principles of Artificial Intelligence

Intelligent Agents. CmpE 540 Principles of Artificial Intelligence CmpE 540 Principles of Artificial Intelligence Intelligent Agents Pınar Yolum pinar.yolum@boun.edu.tr Department of Computer Engineering Boğaziçi University 1 Chapter 2 (Based mostly on the course slides

More information

Credal decision trees in noisy domains

Credal decision trees in noisy domains Credal decision trees in noisy domains Carlos J. Mantas and Joaquín Abellán Department of Computer Science and Artificial Intelligence University of Granada, Granada, Spain {cmantas,jabellan}@decsai.ugr.es

More information

Artificial Intelligence Programming Probability

Artificial Intelligence Programming Probability Artificial Intelligence Programming Probability Chris Brooks Department of Computer Science University of San Francisco Department of Computer Science University of San Francisco p.1/25 17-0: Uncertainty

More information

Learning to Use Episodic Memory

Learning to Use Episodic Memory Learning to Use Episodic Memory Nicholas A. Gorski (ngorski@umich.edu) John E. Laird (laird@umich.edu) Computer Science & Engineering, University of Michigan 2260 Hayward St., Ann Arbor, MI 48109 USA Abstract

More information

Meta-Analysis. Zifei Liu. Biological and Agricultural Engineering

Meta-Analysis. Zifei Liu. Biological and Agricultural Engineering Meta-Analysis Zifei Liu What is a meta-analysis; why perform a metaanalysis? How a meta-analysis work some basic concepts and principles Steps of Meta-analysis Cautions on meta-analysis 2 What is Meta-analysis

More information

Introduction and Historical Background. August 22, 2007

Introduction and Historical Background. August 22, 2007 1 Cognitive Bases of Behavior Introduction and Historical Background August 22, 2007 2 Cognitive Psychology Concerned with full range of psychological processes from sensation to knowledge representation

More information

Probabilistically Estimating Backbones and Variable Bias: Experimental Overview

Probabilistically Estimating Backbones and Variable Bias: Experimental Overview Probabilistically Estimating Backbones and Variable Bias: Experimental Overview Eric I. Hsu, Christian J. Muise, J. Christopher Beck, and Sheila A. McIlraith Department of Computer Science, University

More information

Spatial Orientation Using Map Displays: A Model of the Influence of Target Location

Spatial Orientation Using Map Displays: A Model of the Influence of Target Location Gunzelmann, G., & Anderson, J. R. (2004). Spatial orientation using map displays: A model of the influence of target location. In K. Forbus, D. Gentner, and T. Regier (Eds.), Proceedings of the Twenty-Sixth

More information

Empirical function attribute construction in classification learning

Empirical function attribute construction in classification learning Pre-publication draft of a paper which appeared in the Proceedings of the Seventh Australian Joint Conference on Artificial Intelligence (AI'94), pages 29-36. Singapore: World Scientific Empirical function

More information

Learning Utility for Behavior Acquisition and Intention Inference of Other Agent

Learning Utility for Behavior Acquisition and Intention Inference of Other Agent Learning Utility for Behavior Acquisition and Intention Inference of Other Agent Yasutake Takahashi, Teruyasu Kawamata, and Minoru Asada* Dept. of Adaptive Machine Systems, Graduate School of Engineering,

More information

Memory Prediction Framework for Pattern Recognition: Performance and Suitability of the Bayesian Model of Visual Cortex

Memory Prediction Framework for Pattern Recognition: Performance and Suitability of the Bayesian Model of Visual Cortex Memory Prediction Framework for Pattern Recognition: Performance and Suitability of the Bayesian Model of Visual Cortex Saulius J. Garalevicius Department of Computer and Information Sciences, Temple University

More information

Learning Deterministic Causal Networks from Observational Data

Learning Deterministic Causal Networks from Observational Data Carnegie Mellon University Research Showcase @ CMU Department of Psychology Dietrich College of Humanities and Social Sciences 8-22 Learning Deterministic Causal Networks from Observational Data Ben Deverett

More information

Minimal Change and Maximal Coherence: A Basis for Belief Revision and Reasoning about Actions

Minimal Change and Maximal Coherence: A Basis for Belief Revision and Reasoning about Actions Minimal Change and Maximal Coherence: A Basis for Belief Revision and Reasoning about Actions Anand S. Rao Australian AI Institute Carlton, Vic-3053 Australia Email: anand@aaii.oz.au Abstract The study

More information

Chapter Three: Sampling Methods

Chapter Three: Sampling Methods Chapter Three: Sampling Methods The idea of this chapter is to make sure that you address sampling issues - even though you may be conducting an action research project and your sample is "defined" by

More information

An Empirical Study of Agent Programs

An Empirical Study of Agent Programs An Empirical Study of Agent Programs A Dynamic Blocks World Case Study in GOAL M. Birna van Riemsdijk and Koen V. Hindriks EEMCS, Delft University of Technology, Delft, The Netherlands {m.b.vanriemsdijk,k.v.hindriks}@tudelft.nl

More information

EEL-5840 Elements of {Artificial} Machine Intelligence

EEL-5840 Elements of {Artificial} Machine Intelligence Menu Introduction Syllabus Grading: Last 2 Yrs Class Average 3.55; {3.7 Fall 2012 w/24 students & 3.45 Fall 2013} General Comments Copyright Dr. A. Antonio Arroyo Page 2 vs. Artificial Intelligence? DEF:

More information

Extensions to a Unified Theory of the Cognitive Architecture. Nishant Trivedi

Extensions to a Unified Theory of the Cognitive Architecture. Nishant Trivedi Extensions to a Unified Theory of the Cognitive Architecture by Nishant Trivedi A Thesis Presented in Partial Fulfillment of the Requirements for the Degree Master of Science Approved April 2011 by the

More information

Artificial Intelligence For Homeopathic Remedy Selection

Artificial Intelligence For Homeopathic Remedy Selection Artificial Intelligence For Homeopathic Remedy Selection A. R. Pawar, amrut.pawar@yahoo.co.in, S. N. Kini, snkini@gmail.com, M. R. More mangeshmore88@gmail.com Department of Computer Science and Engineering,

More information

PART - A 1. Define Artificial Intelligence formulated by Haugeland. The exciting new effort to make computers think machines with minds in the full and literal sense. 2. Define Artificial Intelligence

More information

Artificial Intelligence Lecture 7

Artificial Intelligence Lecture 7 Artificial Intelligence Lecture 7 Lecture plan AI in general (ch. 1) Search based AI (ch. 4) search, games, planning, optimization Agents (ch. 8) applied AI techniques in robots, software agents,... Knowledge

More information

P(HYP)=h. P(CON HYP)=p REL. P(CON HYP)=q REP

P(HYP)=h. P(CON HYP)=p REL. P(CON HYP)=q REP BAYESIAN NETWORKS IN PHILOSOPHY STEPHAN HARTMANN, LUC BOVENS 1. Introduction There is a long philosophical tradition of addressing questions in philosophy of science and epistemology by means of the tools

More information

Exploration and Exploitation in Reinforcement Learning

Exploration and Exploitation in Reinforcement Learning Exploration and Exploitation in Reinforcement Learning Melanie Coggan Research supervised by Prof. Doina Precup CRA-W DMP Project at McGill University (2004) 1/18 Introduction A common problem in reinforcement

More information

COMP329 Robotics and Autonomous Systems Lecture 15: Agents and Intentions. Dr Terry R. Payne Department of Computer Science

COMP329 Robotics and Autonomous Systems Lecture 15: Agents and Intentions. Dr Terry R. Payne Department of Computer Science COMP329 Robotics and Autonomous Systems Lecture 15: Agents and Intentions Dr Terry R. Payne Department of Computer Science General control architecture Localisation Environment Model Local Map Position

More information

A Computational Theory of Belief Introspection

A Computational Theory of Belief Introspection A Computational Theory of Belief Introspection Kurt Konolige Artificial Intelligence Center SRI International Menlo Park, California 94025 Abstract Introspection is a general term covering the ability

More information

An Efficient Hybrid Rule Based Inference Engine with Explanation Capability

An Efficient Hybrid Rule Based Inference Engine with Explanation Capability To be published in the Proceedings of the 14th International FLAIRS Conference, Key West, Florida, May 2001. An Efficient Hybrid Rule Based Inference Engine with Explanation Capability Ioannis Hatzilygeroudis,

More information

Learning with Rare Cases and Small Disjuncts

Learning with Rare Cases and Small Disjuncts Appears in Proceedings of the 12 th International Conference on Machine Learning, Morgan Kaufmann, 1995, 558-565. Learning with Rare Cases and Small Disjuncts Gary M. Weiss Rutgers University/AT&T Bell

More information

How Does Analysis of Competing Hypotheses (ACH) Improve Intelligence Analysis?

How Does Analysis of Competing Hypotheses (ACH) Improve Intelligence Analysis? How Does Analysis of Competing Hypotheses (ACH) Improve Intelligence Analysis? Richards J. Heuer, Jr. Version 1.2, October 16, 2005 This document is from a collection of works by Richards J. Heuer, Jr.

More information

Are Retrievals from Long-Term Memory Interruptible?

Are Retrievals from Long-Term Memory Interruptible? Are Retrievals from Long-Term Memory Interruptible? Michael D. Byrne byrne@acm.org Department of Psychology Rice University Houston, TX 77251 Abstract Many simple performance parameters about human memory

More information

Introduction to Medical Computing

Introduction to Medical Computing CS 2124-2125 Introduction to Medical Computing Stephen M. Watt Artificial Intelligence in Medicine Notes by Jacquelyne Forgette, March 2012. University of Western Ontario CS 2125. Stephen M. Watt AIM Artificial

More information

INTRODUCTION TO MACHINE LEARNING. Decision tree learning

INTRODUCTION TO MACHINE LEARNING. Decision tree learning INTRODUCTION TO MACHINE LEARNING Decision tree learning Task of classification Automatically assign class to observations with features Observation: vector of features, with a class Automatically assign

More information

Causal Models Interact with Structure Mapping to Guide Analogical Inference

Causal Models Interact with Structure Mapping to Guide Analogical Inference Causal Models Interact with Structure Mapping to Guide Analogical Inference Hee Seung Lee (heeseung@ucla.edu) Keith J. Holyoak (holyoak@lifesci.ucla.edu) Department of Psychology, University of California,

More information

Models for Inexact Reasoning. Imprecision and Approximate Reasoning. Miguel García Remesal Department of Artificial Intelligence

Models for Inexact Reasoning. Imprecision and Approximate Reasoning. Miguel García Remesal Department of Artificial Intelligence Models for Inexact Reasoning Introduction to Uncertainty, Imprecision and Approximate Reasoning Miguel García Remesal Department of Artificial Intelligence mgremesal@fi.upm.es Uncertainty and Imprecision

More information

Robotics Summary. Made by: Iskaj Janssen

Robotics Summary. Made by: Iskaj Janssen Robotics Summary Made by: Iskaj Janssen Multiagent system: System composed of multiple agents. Five global computing trends: 1. Ubiquity (computers and intelligence are everywhere) 2. Interconnection (networked

More information

Learning to Identify Irrelevant State Variables

Learning to Identify Irrelevant State Variables Learning to Identify Irrelevant State Variables Nicholas K. Jong Department of Computer Sciences University of Texas at Austin Austin, Texas 78712 nkj@cs.utexas.edu Peter Stone Department of Computer Sciences

More information

Implementation of Perception Classification based on BDI Model using Bayesian Classifier

Implementation of Perception Classification based on BDI Model using Bayesian Classifier Implementation of Perception Classification based on BDI Model using Bayesian Classifier Vishwanath Y 1 Murali T S 2 Dr M.V Vijayakumar 3 1 Research Scholar, Dept. of Computer Science & Engineering, Jain

More information

A Brief Introduction to Bayesian Statistics

A Brief Introduction to Bayesian Statistics A Brief Introduction to Statistics David Kaplan Department of Educational Psychology Methods for Social Policy Research and, Washington, DC 2017 1 / 37 The Reverend Thomas Bayes, 1701 1761 2 / 37 Pierre-Simon

More information

and errs as expected. The disadvantage of this approach is that it is time consuming, due to the fact that it is necessary to evaluate all algorithms,

and errs as expected. The disadvantage of this approach is that it is time consuming, due to the fact that it is necessary to evaluate all algorithms, Data transformation and model selection by experimentation and meta-learning Pavel B. Brazdil LIACC, FEP - University of Porto Rua Campo Alegre, 823 4150 Porto, Portugal Email: pbrazdil@ncc.up.pt Research

More information

Finding Information Sources by Model Sharing in Open Multi-Agent Systems 1

Finding Information Sources by Model Sharing in Open Multi-Agent Systems 1 Finding Information Sources by Model Sharing in Open Multi-Agent Systems Jisun Park, K. Suzanne Barber The Laboratory for Intelligent Processes and Systems The University of Texas at Austin 20 E. 24 th

More information

An Interval-Based Representation of Temporal Knowledge

An Interval-Based Representation of Temporal Knowledge An Interval-Based Representation of Temporal Knowledge James F. Allen Department of Computer Science The University of Rochester Rochester, NY 14627 Abstract This paper describes a method for maintaining

More information

Position Paper: How Certain is Recommended Trust-Information?

Position Paper: How Certain is Recommended Trust-Information? Position Paper: How Certain is Recommended Trust-Information? Uwe Roth University of Luxembourg FSTC Campus Kirchberg 6, rue Richard Coudenhove-Kalergi L-1359 Luxembourg uwe.roth@uni.lu ABSTRACT Nowadays

More information

Module 1. Introduction. Version 1 CSE IIT, Kharagpur

Module 1. Introduction. Version 1 CSE IIT, Kharagpur Module 1 Introduction Lesson 2 Introduction to Agent 1.3.1 Introduction to Agents An agent acts in an environment. Percepts Agent Environment Actions An agent perceives its environment through sensors.

More information

STIN2103. Knowledge. engineering expert systems. Wan Hussain Wan Ishak. SOC 2079 Ext.: Url:

STIN2103. Knowledge. engineering expert systems. Wan Hussain Wan Ishak. SOC 2079 Ext.: Url: & Knowledge STIN2103 engineering expert systems Wan Hussain Wan Ishak SOC 2079 Ext.: 4786 Email: hussain@uum.edu.my Url: http://www.wanhussain.com Outline Knowledge Representation Types of knowledge Knowledge

More information

15.053x. OpenSolver (http://opensolver.org/)

15.053x. OpenSolver (http://opensolver.org/) 15.053x OpenSolver (http://opensolver.org/) 1 Table of Contents Introduction to OpenSolver slides 3-4 Example 1: Diet Problem, Set-Up slides 5-11 Example 1: Diet Problem, Dialog Box slides 12-17 Example

More information

Belief Revision Controlled by Meta-abduction

Belief Revision Controlled by Meta-abduction Belief Revision Controlled by Meta-abduction VIVEK BHARATHAN, Laboratory for Artificial Intelligence Research. Department of Computer Science and Engineering. Ohio State University. Columbus, Ohio (USA).

More information

Unifying Data-Directed and Goal-Directed Control: An Example and Experiments

Unifying Data-Directed and Goal-Directed Control: An Example and Experiments Unifying Data-Directed and Goal-Directed Control: An Example and Experiments Daniel D. Corkill, Victor R. Lesser, and Eva Hudlická Department of Computer and Information Science University of Massachusetts

More information

LEARNING. Learning. Type of Learning Experiences Related Factors

LEARNING. Learning. Type of Learning Experiences Related Factors LEARNING DEFINITION: Learning can be defined as any relatively permanent change in behavior or modification in behavior or behavior potentials that occur as a result of practice or experience. According

More information

Foundations of AI. 10. Knowledge Representation: Modeling with Logic. Concepts, Actions, Time, & All the Rest

Foundations of AI. 10. Knowledge Representation: Modeling with Logic. Concepts, Actions, Time, & All the Rest Foundations of AI 10. Knowledge Representation: Modeling with Logic Concepts, Actions, Time, & All the Rest Wolfram Burgard, Andreas Karwath, Bernhard Nebel, and Martin Riedmiller 10/1 Contents Knowledge

More information

International Journal of Software and Web Sciences (IJSWS)

International Journal of Software and Web Sciences (IJSWS) International Association of Scientific Innovation and Research (IASIR) (An Association Unifying the Sciences, Engineering, and Applied Research) ISSN (Print): 2279-0063 ISSN (Online): 2279-0071 International

More information

A Scoring Policy for Simulated Soccer Agents Using Reinforcement Learning

A Scoring Policy for Simulated Soccer Agents Using Reinforcement Learning A Scoring Policy for Simulated Soccer Agents Using Reinforcement Learning Azam Rabiee Computer Science and Engineering Isfahan University, Isfahan, Iran azamrabiei@yahoo.com Nasser Ghasem-Aghaee Computer

More information

Survival Skills for Researchers. Study Design

Survival Skills for Researchers. Study Design Survival Skills for Researchers Study Design Typical Process in Research Design study Collect information Generate hypotheses Analyze & interpret findings Develop tentative new theories Purpose What is

More information

Tackling Random Blind Spots with Strategy-Driven Stimulus Generation

Tackling Random Blind Spots with Strategy-Driven Stimulus Generation Tackling Random Blind Spots with Strategy-Driven Stimulus Generation Matthew Ballance Mentor Graphics Corporation Design Verification Technology Division Wilsonville, Oregon matt_ballance@mentor.com Abstract

More information

BAYESIAN NETWORK FOR FAULT DIAGNOSIS

BAYESIAN NETWORK FOR FAULT DIAGNOSIS BAYESIAN NETWOK FO FAULT DIAGNOSIS C.H. Lo, Y.K. Wong and A.B. ad Department of Electrical Engineering, The Hong Kong Polytechnic University Hung Hom, Kowloon, Hong Kong Fax: +852 2330 544 Email: eechlo@inet.polyu.edu.hk,

More information

How to Create Better Performing Bayesian Networks: A Heuristic Approach for Variable Selection

How to Create Better Performing Bayesian Networks: A Heuristic Approach for Variable Selection How to Create Better Performing Bayesian Networks: A Heuristic Approach for Variable Selection Esma Nur Cinicioglu * and Gülseren Büyükuğur Istanbul University, School of Business, Quantitative Methods

More information

CS148 - Building Intelligent Robots Lecture 5: Autonomus Control Architectures. Instructor: Chad Jenkins (cjenkins)

CS148 - Building Intelligent Robots Lecture 5: Autonomus Control Architectures. Instructor: Chad Jenkins (cjenkins) Lecture 5 Control Architectures Slide 1 CS148 - Building Intelligent Robots Lecture 5: Autonomus Control Architectures Instructor: Chad Jenkins (cjenkins) Lecture 5 Control Architectures Slide 2 Administrivia

More information

Supplementary notes for lecture 8: Computational modeling of cognitive development

Supplementary notes for lecture 8: Computational modeling of cognitive development Supplementary notes for lecture 8: Computational modeling of cognitive development Slide 1 Why computational modeling is important for studying cognitive development. Let s think about how to study the

More information

Knowledge Representation defined. Ontological Engineering. The upper ontology of the world. Knowledge Representation

Knowledge Representation defined. Ontological Engineering. The upper ontology of the world. Knowledge Representation 3 Knowledge Representation defined Knowledge Representation (Based on slides by Tom Lenaerts) Lars Bungum (Ph.D. stip.) Department of Computer & Information Science How to represent facts about the world

More information

Strategic behaviour-based reasoning with. S Linkoping, Sweden. Abstract. Behaviour-based agents rely on current perceptions for behaviour

Strategic behaviour-based reasoning with. S Linkoping, Sweden. Abstract. Behaviour-based agents rely on current perceptions for behaviour Strategic behaviour-based reasoning with dynamic, partial information James Westendorp y? Paul Scerri z Lawrence Cavedon y y Dept of Computer Science, RMIT GPO Box 2476V, Melbourne 3001, Australia fjhw,

More information

The Power of Feedback

The Power of Feedback The Power of Feedback 35 Principles for Turning Feedback from Others into Personal and Professional Change By Joseph R. Folkman The Big Idea The process of review and feedback is common in most organizations.

More information

[1] provides a philosophical introduction to the subject. Simon [21] discusses numerous topics in economics; see [2] for a broad economic survey.

[1] provides a philosophical introduction to the subject. Simon [21] discusses numerous topics in economics; see [2] for a broad economic survey. Draft of an article to appear in The MIT Encyclopedia of the Cognitive Sciences (Rob Wilson and Frank Kiel, editors), Cambridge, Massachusetts: MIT Press, 1997. Copyright c 1997 Jon Doyle. All rights reserved

More information

A Bayesian Approach to Tackling Hard Computational Challenges

A Bayesian Approach to Tackling Hard Computational Challenges A Bayesian Approach to Tackling Hard Computational Challenges Eric Horvitz Microsoft Research Joint work with: Y. Ruan, C. Gomes, H. Kautz, B. Selman, D. Chickering MS Research University of Washington

More information

Evaluation Models STUDIES OF DIAGNOSTIC EFFICIENCY

Evaluation Models STUDIES OF DIAGNOSTIC EFFICIENCY 2. Evaluation Model 2 Evaluation Models To understand the strengths and weaknesses of evaluation, one must keep in mind its fundamental purpose: to inform those who make decisions. The inferences drawn

More information

Agent-Based Systems. Agent-Based Systems. Michael Rovatsos. Lecture 5 Reactive and Hybrid Agent Architectures 1 / 19

Agent-Based Systems. Agent-Based Systems. Michael Rovatsos. Lecture 5 Reactive and Hybrid Agent Architectures 1 / 19 Agent-Based Systems Michael Rovatsos mrovatso@inf.ed.ac.uk Lecture 5 Reactive and Hybrid Agent Architectures 1 / 19 Where are we? Last time... Practical reasoning agents The BDI architecture Intentions

More information

Utility Maximization and Bounds on Human Information Processing

Utility Maximization and Bounds on Human Information Processing Topics in Cognitive Science (2014) 1 6 Copyright 2014 Cognitive Science Society, Inc. All rights reserved. ISSN:1756-8757 print / 1756-8765 online DOI: 10.1111/tops.12089 Utility Maximization and Bounds

More information

Emotions in Intelligent Agents

Emotions in Intelligent Agents From: FLAIRS-02 Proceedings. Copyright 2002, AAAI (www.aaai.org). All rights reserved. Emotions in Intelligent Agents N Parameswaran School of Computer Science and Engineering University of New South Wales

More information

Recognizing Ambiguity

Recognizing Ambiguity Recognizing Ambiguity How Lack of Information Scares Us Mark Clements Columbia University I. Abstract In this paper, I will examine two different approaches to an experimental decision problem posed by

More information

Objectives. Quantifying the quality of hypothesis tests. Type I and II errors. Power of a test. Cautions about significance tests

Objectives. Quantifying the quality of hypothesis tests. Type I and II errors. Power of a test. Cautions about significance tests Objectives Quantifying the quality of hypothesis tests Type I and II errors Power of a test Cautions about significance tests Designing Experiments based on power Evaluating a testing procedure The testing

More information

Dynamic Rule-based Agent

Dynamic Rule-based Agent International Journal of Engineering Research and Technology. ISSN 0974-3154 Volume 11, Number 4 (2018), pp. 605-613 International Research Publication House http://www.irphouse.com Dynamic Rule-based

More information

Inductive Learning of Simple Diagnostic Scores

Inductive Learning of Simple Diagnostic Scores Inductive Learning of Simple Diagnostic Scores Martin Atzmueller, Joachim Baumeister, and Frank Puppe University of Würzburg, 97074 Würzburg, Germany Department of Computer Science Phone: +49 931 888-6746,

More information

COHERENCE: THE PRICE IS RIGHT

COHERENCE: THE PRICE IS RIGHT The Southern Journal of Philosophy Volume 50, Issue 1 March 2012 COHERENCE: THE PRICE IS RIGHT Paul Thagard abstract: This article is a response to Elijah Millgram s argument that my characterization of

More information

What Is Science? Lesson Overview. Lesson Overview. 1.1 What Is Science?

What Is Science? Lesson Overview. Lesson Overview. 1.1 What Is Science? Lesson Overview 1.1 What Science Is and Is Not What are the goals of science? One goal of science is to provide natural explanations for events in the natural world. Science also aims to use those explanations

More information

How do you design an intelligent agent?

How do you design an intelligent agent? Intelligent Agents How do you design an intelligent agent? Definition: An intelligent agent perceives its environment via sensors and acts rationally upon that environment with its effectors. A discrete

More information