Ingo Thon's Website

Using machine learning techniques for planning is getting increasingly more important in recent years. Various aspects of action models can be induced from data and then exploited for planning. For probabilistic planning, natural candidates are learning of action effects and their probabilities. For expressive formalisms such as PPDDL, this is a difficult prob- lem since they can introduce easily a hidden data problem; the fact that multiple action outcomes may have generated the same experienced state transitions in the data. Furthermore the action effects might be factored such that this prob- lem requires solving a constraint satisfaction problem within an expectation maximization scheme. In this paper we outline how to utilize recent techniques from the field of statistical relational learning for this problem. More specifically, we show how techniques developed for the CPT-L model of relational probabilistic sequences can be applied to the problem of learning probabilities in a PPDDL model. A CPT-L model concisely specify a Markov chain over arbitrary numbers of objects in the domain and simultaneous applications of multiple actions. The use of efficient BDD-style representations allows for fast and efficient learning in such domains. Even efficient online learning is possible as we will show in this paper. We relate to other learning approaches for similar domains and highlight the opportunities for incorporating our approach into architectures that can plan, execute the plan, and learn from the outcomes, in an online and incremental fashion.

One of the current challenges in artificial intelligence is modeling dynamic environments that change due to the actions or activities undertaken by people or agents. The task of inferring hidden states, e.g. the activities or intentions of people, based on observations is called filtering. Standard probabilistic models such as Dynamic Bayesian Networks are able to solve this task efficiently using approximative methods such as particle filters. However, these models do not support logical or relational representations. The key contribution of this paper is the upgrade of a particle filter algorithm for use with a probabilistic logical representation through the definition of a proposal distribution. The performance of the algorithm depends largely on how well this distribution fits the target distribution. We adopt the idea of logical compilation into Binary Decision Diagrams for sampling. This allows us to use the optimal proposal distribution which is normally prohibitively slow.

Understanding complex, dynamic scenes of real-world activities from low-level sensor data is of central importance for intelligent systems. The main difficulty lies in the fact that complex scenes are best described in high-level, logical formalisms, while sensor data usually consists of many low-level features. We first propose a method to obtain a logical representation of real-world, dynamic scenes based on input video stream solely. We focus on representing the video data using probabilistic relational sequences as a natural way to incorporate sensor uncertainty. They allow us to work with structured terms, and in addition they capture the inherent uncertainty of object detection. Further on, we employ r-grams as the probabilistic logical learning model for this application. In a first step we use r-grams in a simple setting and we show their viability in card games. We also show how r-grams can be upgraded to deal with uncertain observations.

We report on two exercises in modeling, in ference and learning with seven statistical relational learning systems and use this as a basis for a simple and preliminary comparison between these systems

Being able to understand complex dynamic scenes of real-world activities from low-level sensor data is of central importance for intelligent systems. The main difficulty lies in the fact that complex scenes are best described in high-level, logical formalisms, whereas sensor data usually consists of many low-level feature values. In this work, we consider the problem of learning high-level, logical descriptions of dynamic scenes based on input video stream solely. In order to learn such general patterns, two important problems must be tackled and their solutions combined: obtaining high-level, logical representations from video data and learning probabilistic logical models of dynamic scenes. This setting opens new research directions. We focus on representing the video data using probabilistic relational sequences as a natural way to incorporate sensor information in real-world tasks. They allow to work with structured terms, but in addition they capture the inherent uncertainty of object detection. Further on, we employ relational sequence learning methods for this type of video representation. We propose table-top and card games as application domain.

This paper reports on our work towards the development of a probabilistic logic programming environment intended as a target language in which other probabilistic languages can be compiled, thereby contributing to the digestion of the "alphabet soup".

Artificial intelligence aims at developing agents that learn and act in complex environments. Realistic environments typically feature a variable number of objects, relations amongst them, and non-deterministic transition behavior. Standard probabilistic sequence models provide efficient inference and learning techniques, but typically cannot fully capture the relational complexity. On the other hand, statistical relational learning techniques are often too inefficient. In this paper, we present a simple model that occupies an intermediate position in this expressiveness/efficiency trade-off. Based on CP-logic, an expressive probabilistic logic for modeling causality. but specialized to represent a probability distribution over sequences of relational state descriptions, and employing a Markov assumption, inference and learning become more tractable and effective.

Artificial intelligence aims at developing agents that learn and act in complex environments. Realistic environments typically feature a variable number of objects, relations amongst them, and non-deterministic transition behavior. Standard probabilistic sequence models provide efficient inference and learning techniques, but typically cannot fully capture the relational complexity. On the other hand, statistical relational learning techniques are often too inefficient. In this paper, we present a simple model that occupies an intermediate position in this expressiveness/efficiency trade-off. It is based on CP-logic, an expressive probabilistic logic for modeling causality. However, by specializing CP-logic to represent a probability distribution over sequences of relational state descriptions, and employing a Markov assumption, inference and learning become more tractable and effective. We show that the resulting model is able to handle probabilistic relational domains with a substantial number of objects and relations.

Agents that learn and act in real-world environments have to cope with both complex state descriptions and non-deterministic transition behavior of the world. Standard statistical relational learning techniques can capture this complexity, but are often inefficient. We present a simple probabilistic model for such environments based on CP-Logic. Efficiency is maintained by restriction to a fully observable setting and the use of efficient inference algorithms based on binary decision diagrams.

Agents that learn and act in real-world environments have to cope with both complex state descriptions and non-deterministic transition behavior of the world. Standard statistical relational learning techniques can capture this complexity, but are often inefficient. We present a simple probabilistic model for such environments based on CP-Logic. efficiency is maintained by restriction to a fully observable setting.

The ability to recognize human activities from sensory information is essential for developing the next generation of smart devices. Many human activity recognition tasks are - from a machine learning perspective - quite similar to tagging tasks in natural language processing. Motivated by this similarity, we develop a relational transformation-based tagging system based on inductive logic programming principles, which is able to cope with expressive relational representations as well as a background theory. The approach is experimentally evaluated on two activity recognition tasks and an information extraction task, and compared to Hidden Markov Models, one of the most popular and successful approaches for tagging.

Several promising variants of hidden Markov models (HMMs) have recently been developed to efficiently deal with large state and observation spaces and relational structure. Many application domains, however, have an apriori componential structure such as parts in musical scores. In this case, exact inference within relational HMMs still grows exponentially in the number of components. In this paper, we propose to approximate the complex joint relational HMM with a simpler, distributed one: k relational hidden chains over n states, one for each component. Then, we iteratively perform inference for each chain given fixed values for the other chains until convergence. Due to this structured mean field approximation, the effective size of the hidden state space collapses from O(n to the power of k ) to O(k * n).