Howe, A.E., Cohen, P.R.: Understanding planner behavior. AIJ 76 (1995) 125--166  Home/Search   Document Details and Download   Summary   Related Articles   Check

....In this case, a subsequence Pass(P layer 2 ; P layer 3 ) Dribble(P layer 3 ) would not have been a useful sequence to discover it reveals nothing about the coordinated behavior of the team. We present a technique for processing such a stream of events, building on earlier work by Howe and Cohen [8], and extending it in novel ways. The technique we present rejects such false sequences, by uncovering sequential statistical dependencies between atomic observed behaviors of agents. We demonstrate the e cacy of our approach in learning sequential behavior of several RoboCup simulation teams from ....

....statistical dependencies between atomic observed behaviors of agents. We demonstrate the e cacy of our approach in learning sequential behavior of several RoboCup simulation teams from observations, and compare the results of using frequency counts [9] and statistical dependency detection [8], two important sequencemining techniques, in identifying important sequences. We show that while the two techniques are both signi cantly better than random selection, statistical dependency detection has important bene ts in being able to distinguish causal dependencies between observed ....

[Article contains additional citation context not shown here]

Howe, A.E., Cohen, P.R.: Understanding planner behavior. Articial Intelligence 76 (1995) 125166

....to it. In this case, a subsequence Pass(Player2, Player3) Dribble(Player3) would not have been a useful sequence to discover it reveals nothing about the coordinated behavior of the team. We present a technique for processing such a stream of events, building on earlier work by Howe and Cohen [8], and extending it in novel ways. The tech nique we present rejects such false sequences, by uncovering sequential statistical dependencies between atomic observed behaviors of agents. We demonstrate the efficacy of our approach in learning sequential behavior of several RoboCup simulation teams ....

....statistical dependencies between atomic observed behaviors of agents. We demonstrate the efficacy of our approach in learning sequential behavior of several RoboCup simulation teams from observations, and compare the results of using frequency counts [9] and statistical dependency detection [8], two important sequence mining techniques, in identifying important sequences. We show that while the two techniques are both significantly better than random selection, statistical dependency detection has important benefits in being able to distinguish causal dependencies between observed ....

[Article contains additional citation context not shown here]

Howe, A.E., Cohen, P.R.: Understanding planner behavior. Artificial Intelligence 76 (1995) 125-166

....vs. 8 10) since the incorrect plan provided context and the oracle identified the o#ending plan elements. Automated diagnosis could reduce these e#orts, especially for determining why the planner failed to generate a plan. There has been some work in this area that could be applied or extended. Howe (Howe Cohen 1995) performed statistical analyses of the planner trace to identify applications of repair operators to states that were strongly correlated with failures. Chien (Chien 1998) allowed the planner to generate a plan, when it was otherwise unable to, by ignoring problematic constraints. Analysts were ....

Howe, A. E., and Cohen, P. R. 1995. Understanding planner behavior. Artificial Intelligence 76(2):125-- 166.

.... short pieces of the process execution trace (called subsequences) and use statistics to find significant subsequences which have unusually high or low frequency of occurrence (dependencies) They are part of a family of algorithms called Dependency Detection, or in short, DD (described in [7, 8, 9, 10, 11]) The resulting dependencies are sometimes referred to as snapshots. The algorithm I developed uses the snapshots obtained from members of the DD algorithm family, and integrates them using heuristic techniques. The result is an overview model, which in fact is a state transition diagram (STD) ....

.... it is often referred to as a test of heterogeneity, i.e. it tests whether the ratios of cell frequencies in the two rows are heterogeneous (the more heterogeneous these rows are, the more significant is our dependency) The use of the G test in snapshot DD is also preferred by Howe and Cohen [9] for the subsumption snapshot DD algorithm (see subsection 2.1.2) because unlike Chi Square it is fully additive (as described in [23] 2.1.2 Types of Snapshot Dependencies The simplest form of snapshot DD is called absolute order DD. Subsequences of length n are counted and analyzed. The ....

[Article contains additional citation context not shown here]

Adele E. Howe, Paul R. Cohen. "Understanding Planner Behavior", Artificial Intelligence, vol. 76(1-2), pp. 125-166, 1995.

....tree [4] for the mining engine, to learn separate models for each arc in the path. In our domain this would correspond to learning rules for each route in the evacuation domain. However, our goal is different in that we are trying to learn long sequences of events that cause plan failure. In [13] a methodology called dependency interpretation is presented, that uses statistical dependency detection to identify interesting (unusually frequent or infrequent) patterns in plan execution traces. They then interpret the patterns using a weak model of the planner s interaction with its ....

A. E. Howe and P. R. Cohen. Understanding planner behavior. J. Artificial Intelligence, 76(1):125--166, 1995.

....include a mechanism for incrementally stepping through plan generation and viewing the plan being developed (e.g. 18,5] Both of these approaches emphasize the plan generation phase, assuming that execution should proceed as planned. Execution was the focus of Failure Recovery Analysis (FRA) [7,11]. FRA is a methodology for debugging some canonical bug types in the Phoenix planner by statistically analyzing traces of failure recovery. We focused on statistical analysis because the subtle differences in context, the reliance on occasional stochastic decision making and the time periods in ....

....on the right side of Figure 2 resulted. This diagram collapses the four states into two, making it more compact, but also removing one possible transition (AB CD) 3 What We Have Learned So Far Snapshot dependencies have been used to support debugging failure recovery in the Phoenix planner [11] and to identify search control problems in UCPOP[17] However, the dependencies 4 (E C D) 155 1 19 823) B E C) 153 4 95 746) C D A) 87 81 7 823) D A B) 86 2 235 675) D B E) 78 1 83 836) C D B) 75 93 246 584) A B B) 74 18 247 659) B B C) 74 1 174 749) B C B) 74 14 247 663) C ....

Adele E. Howe and Paul R. Cohen. Understanding planner behavior. Artificial Intelligence, 76(1-2):125--166, 1995.

.... 15,000 Assistant 86 76 ; CN2 82 ; Bayes 83 NetTalk 70.11 50,000 NetTalk 77 Monks 2 79.17 5,000 CN2 69.0 ; ID3 69.1 ; back prop 100 Mushroom 99.49 30,000 GINI 98.6 ; Info Gain 98.6 ; C4 100 Soybean 13.83 20,000 IWN 97.1 Thyroid 95.46 20,000 Waveform 40 73.02 15,000 Nearest neighbor 38 ; CART 72 ; Bayes 86 TABLE 18.1. Performance of msdd as a feature based classifier on thirteen datasets from the UC Irvine collection. 18.4.2 Pathology Prediction We applied msdd to the task of predicting pathologies in a simulated shipping network called TransSim. When several ships attempt ....

....fluctuations in the state of the simulation. This behavior is beneficial when we view disruption to the original schedule 192 Tim Oates, Matthew D. Schmill, Dawn E. Gregory and Paul R. Cohen as a cost that we want to minimize. Cost Demon Mean Rule Mean p Value PP 184.2 94.6 0.0001 CT 2289. 3 2377.9 0.0689 IC 1149.8 1202.1 0.1844 QL 637.7 640.5 0.9177 SD 131.1 141.6 0.0019 SU 188.8 202.2 0.3475 SM 21.6 9.2 0.0001 TABLE 18.2. Comparison of simulation costs using demon and msdd rules for pathology prediction. This experiment suggests that msdd can discover indicators of pathological ....

[Article contains additional citation context not shown here]

Howe, Adele E. and Cohen, Paul R. Understanding Planner Behavior. To appear in AI Journal, Winter 1995.

....) minimizes the expected cost of a sequence of methods [50] A key assumption is that P r(M j j S i ) is independent of the order in which methods are tried. We showed that in Phoenix, at least, this assumption is wrong [37] which led to some techniques for detecting interactions among methods [38, 39]. These early results all involve simple models in which terms represent features of the Phoenix environment (e.g. the probability of a fire starting) or agent architecture (e.g. the cost of trying a method) or tasks (e.g. the rate of fire growth) The models say nothing about agent behavior, ....

....specifically, a failure F 1 , the method that repaired it M 1 , and the subsequent failure F 2 . The patterns of dependency were quite complex; for example, for some F 1 , the application of M 1 decreased the incidence of F 2 , but for other initial failures, M 1 increased the incidence of F 2 [39]. Dependency detection can be generalized to multiple execution traces. The Phoenix fireboss, for example, controls several bulldozers, each of which performs a sequence of actions. Is it possible to find statistical dependencies between the actions taken by different bulldozers at different ....

Adele E. Howe and Paul R. Cohen. Understanding planner behavior. AI Journal, 1995. Forthcoming.

....(DD) algorithms, is a technique for identifying situations that lead to failure in AI planning systems. The procedure, termed failure recovery analysis (FRA) models a program as a single stream (the execution trace) and uses a statistical test 1 to locate contributors to plan failure. [4] The benefit of using FRA as a debugging tool for planning systems is due largely to its generality. Basing its diagnosis solely on observed patterns in execution traces allows the dependency detection algorithm to operate with little domain knowledge and only a weak model of how the actual system ....

Adele E. Howe and Paul R. Cohen. Understanding planner behavior. To appear in AI Journal, Winter 1995.

....two AI systems. 1.1 Applications and Techniques for Event Sequence Modeling Our primary application for event sequence modeling is debugging. Our method for generating snapshot models, called dependency detection (DD) has been used to support debugging failure recovery in the Phoenix planner [9] and to identify search control problems in UCPOP[16] Snapshot DD methods discover unusually frequently or infrequently co occurring sequences of events (called dependencies) using contingency table analysis [8] see Section 2.1) Mannila et al. 12] find serial (i.e. a strict ordering of ....

Adele E. Howe and Paul R. Cohen. Understanding planner behavior. Artificial Intelligence, 76(1-2):125--166, 1995.

....the agent and its goal at each point in time. The boundary of the shaded region is an envelope, specifically a slack time envelope, for this task [Cohen92] When the agent crosses this boundary, the envelope violates, or signals a failure, which typically requires some modification to a plan [Howe93, Howe95]. The slacktime aspect of the envelope refers to an initial period of time during which no failure predictions are made. The agent is thus allowed to fall slightly behind at first, under the assumption that it can make up for lost time later. The slack time envelope is based on the determination ....

Howe, Adele E. and Cohen, P.R., 1995. Understanding Planner Behavior. AI Journal. To appear.

....difficult to predict and problems difficult to replicate. Program actions often interact in unforeseen and deleterious ways. We employ a technique we call dependency detection, analyzing program execution traces with a statistical filter to find significant dependencies among interacting actions [9]. Once identified, these dependencies can be examined more carefully to find and fix the unforeseen interactions that often cause them. We have successfully employed dependency detection to identify and debug such interactions in the Phoenix planner, using execution traces that consist of a single ....

Adele E. Howe and Paul R. Cohen. Understanding planner behavior. Artificial Intelligence. To appear.

....difficult to predict and problems difficult to replicate. Program actions often interact in unforeseen and deleterious ways. We employ a technique we call dependency detection, analyzing program execution traces with a statistical filter to find significant dependencies among interacting actions [10, 9, 12]. Causal Induction Having explored the data and or identified dependencies among interacting factors, the user next tries to build a predictive model of the program s behavior. We would like for such a model to tell us how to change the program to improve or modify its behavior. This requires ....

Adele E. Howe and Paul R. Cohen. Understanding planner behavior. Artificial Intelligence. To appear.

.... Bayes 74 Lymphography 78.16 15,000 Assistant 86 76 ; CN2 82 ; Bayes 83 NetTalk 70.11 50,000 NetTalk 77 Monks 2 79.17 5,000 CN2 69.0 ; ID3 69.1 ; back prop 100 Mushroom 99.49 30,000 GINI 98.6 ; Info Gain 98.6 ; C4 100 Soybean 13.83 20,000 IWN 97.1 Thyroid 95.46 20,000 Waveform 40 73.02 15,000 Nearest neighbor 38 ; CART 72 ; Bayes 86 TABLE 1.1. Performance of msdd as a feature based classifier on thirteen datasets from the UC Irvine collection. from materializing. Using the demon as an oracle, we gathered data from a single run of the simulator and used msdd to generate ....

.... 15,000 Assistant 86 76 ; CN2 82 ; Bayes 83 NetTalk 70.11 50,000 NetTalk 77 Monks 2 79.17 5,000 CN2 69.0 ; ID3 69.1 ; back prop 100 Mushroom 99.49 30,000 GINI 98.6 ; Info Gain 98.6 ; C4 100 Soybean 13.83 20,000 IWN 97.1 Thyroid 95.46 20,000 Waveform 40 73.02 15,000 Nearest neighbor 38 ; CART 72 ; Bayes 86 TABLE 1.1. Performance of msdd as a feature based classifier on thirteen datasets from the UC Irvine collection. from materializing. Using the demon as an oracle, we gathered data from a single run of the simulator and used msdd to generate rules to predict bottlenecks. To ....

[Article contains additional citation context not shown here]

Howe, Adele E. and Cohen, Paul R. Understanding Planner Behavior. To appear in AI Journal, Winter 1995.

....For a more practical demonstration of our method, we constructed behavioral models for two systems for planning and control: Phoenix and RARS. 5. 1 The Phoenix Planner Dependency Detection was originally motivated by the difficulties encountered in debugging the Phoenix planning system [11]. The Phoenix system encompasses a simulation of forest fire fighting in Yellowstone National Park and the agents that operate within it [6] Typically, we configured the simulation to include about 10 agents working together over 60 100 hours of simulated time. Decisions about what actions to ....

....between some of the recovery methods and later failures. By searching for static and dynamic connections between the plan actions involved, we identified bugs (e.g. variables calculated incorrectly and improper plan synchronization) using a methodology called Failure Recovery Analysis[11]. However, these bugs still involved fairly simple interactions local to just a few events. Some of the more pernicious failures were included in many dependencies, making it difficult to determine what was causing the problem. We used the original execution data 1 on which the recovery code ....

A. E. Howe and P. R. Cohen. Understanding planner behavior. Artificial Intelligence, 76(1-2):125--166, 1995.

....for ongoing processes, called fluents, and describe an algorithm for finding fluents and associations among them in multiple streams of data. Finding Structure in a Single Stream The first algorithm was developed by Adele Howe and Paul Cohen for finding dependencies in a single stream (Howe and Cohen, 1995). Consider the Wave stream, above. It contains two tokens, W and O, which apparently are not distributed uniformly. A simple G test on a contingency table tells us that W follows itself more often than we d expect by chance under the uniform distribution hypothesis. The contingency table is shown ....

....occurrences of W are not independent of immediately prior occurrences. It would be easy to introduce a lag into the analysis to find dependencies between one token and another after, say, five time steps, and Howe has designed an adaptive algorithm to find the most predictive lag for dependencies (Howe, 1995). Note also that the average length of a run of W s is just the first row margin (13) divided by the cell 2 count, that is, 13 2 = 7.5. Although this technique is very simple, Howe used it to find dependencies between different events in the execution traces of complex computer programs, ....

[Article contains additional citation context not shown here]

Howe, A.E. and P.R. Cohen. (1995) Understanding Planner Behavior. To appear in AI Journal.

....that is, neither learns dependencies between sequences of multitokens. We will return to this issue later. The multi stream dependency detection algorithm (msdd) is an extension of Adele Howe s algorithm for finding dependencies between tokens in a single stream (Oates, Gregory and Cohen, 1995; Howe and Cohen, 1995). We will describe Howe s algorithm first. Let (p; s; ffi ) denote a dependency. Each dependency rule says that when the precursor token, p, occurs at time step i in the stream, the successor token, s, will occur at time step i ffi in the stream with some probability. When this probability is ....

A. E. Howe and P. R. Cohen. Understanding Planner Behavior. To appear in AI Journal, 1995.

....structure in the data, regardless of whether an accurate set of target concepts exists. 5 Related Work The research reported in this paper grew out of the work of Howe and Cohen on finding dependencies between events in execution traces (a single stream of data) generated by the Phoenix planner (Howe Cohen 1995). Dependencies between planner failures, failure recovery actions, and subsequent failures were combined with a weak model of the planner to automate analysis and debugging of recovery mechanisms. msdd extends the dependency detection portion of that work by considering a significantly more ....

Howe, A. E., and Cohen, P. R. 1995. Understanding planner behavior. Artificial Intelligence 76(1--2):125-- 166.

....and for collecting the RARS data. stepping through plan generation and viewing the plan being developed (e.g. 18, 5] Both of these approaches emphasize the plan generation phase, assuming that execution should proceed as planned. Execution was the focus of Failure Recovery Analysis (FRA) [7, 11]. FRA is a methodology for debugging some canonical bug types in the Phoenix planner by statistically analyzing traces of failure recovery. We focused on statistical analysis because the subtle differences in context, the reliance on occasional stochastic decision making and the time periods in ....

.... (87 81 7 823) C D B) 75 93 246 584) B C B) 74 14 247 663) Table 2: Length 3 dependencies (sequence with contingency table) collected for synthetic five token model 3 What We Have Learned So Far Snapshot dependencies have been used to support debugging failure recovery in the Phoenix planner [11] and to identify search control problems in UCPOP[17] However, the dependencies were rather limited in their temporal scope and complexity; in some cases, especially with Phoenix, the short sequences led us to make changes that caused other problems, changes that would have been different had we ....

Adele E. Howe and Paul R. Cohen. Understanding planner behavior. Artificial Intelligence, 76(1-2):125--166, 1995.

No context found.

Howe, A.E., Cohen, P.R.: Understanding planner behavior. AIJ 76 (1995) 125--166

No context found.

A. E. Howe and P. R. Cohen. Understanding planner behavior. AIJ, 76(1--2):125--166, 1995.

No context found.

A. E. Howe and P. R. Cohen. Understanding planner behavior. AIJ, 76(1--2):125--166, 1995.

Online articles have much greater impact   More about CiteSeer.IST   Add search form to your site   Submit documents   Feedback  

CiteSeer.IST - Copyright Penn State and NEC