AAAI Doctor consortium summary (day one)

Kiri, the chair, gave a welcome speech and 8 PhD students presented their theses work focusing on a leaning and planning.  (thanks Mykel for helping me revising this post)

Jennifer Neville (UMass) -- Structure Learning for Statistical Relational Models

Motivation: latent groups could be detected from graph structure as well as node attributes. Subjects: graph analysis, clustering Issues: (i) Partitioning a network into groups using both link structure analysis and node properties clustering (using EM); (ii) utilizing group assignment for better attribute-value predication and unbiased feature selection. Comments: Groups, as mentioned in this talk, are disjoint; however, further work might involve instances where a node may belong to multiple groups. This problem can be viewed as a clustering problem which uses features from node attributes as well as graph structure information.

Shimon Whiteson (UTexas Austin) --Improving Reinforcement Learning Function Approximators via Neuroevolution Motivation: adaptive scheduling policy can be learned using neuroevolution. Subjects: function approximation, reinforcement learning, evolutionary neural network. Issues: (i) the transition matrix (state-action table) can be huge, and a neural network (NN) based Function Approximator (FA) is a compact alternative; (ii) NEAT+Q, which evolves both the weights and the topology of the NN using NEAT and Q learning, is helpful. Experiments show that Darwinian evolution achieves the best performance compared to Lamarkian evolution, randomized strategies, etc. Comments: Many are skeptical about whether evolving NN could be used for online scheduling since it takes a lot of computational time. In order to evaluate the significance of improvement, the optimal and worse cases are needed.

Bhaskara Marthi (UC Berkeley) -- Discourse Factors in Multi-Document Summarization Motivation: Decomposition is needed for planning asynchronous tasks and many joint choices in a giant state space. Subjects: planning, reinforcement learning, optimization Issues: (i) concurrent ALisp and coordination graphs are used for concurrent hierarchical task decomposition; (ii) decompose Q function w.r.t. subroutines, and that requires reward decomposition which is hard. Comments: Decomposing Q function for the reward of "exiting a subroutine" is hard since there could be thousands of ways to exit.

Trey Smith (CMU) --Rover Science Autonomy: Probabilistic Planning for Science-Aware Exploration Motivation: Discover scientifically interesting objects in extreme environments (e.g., Mars). Subjects: Planning Issues: (i) planning navigation -- with maximum coverage over a spatial extent; (ii) selective sampling — variety is preferred to large amount of copies of samples.

---

Marie desJardins (mentor from UMBC) Prepare talks for different context:

1. Job talk. Try to amuse people with your work and how complex it is.
2. Doctor consortium talk. Present technical details, expose your strength and weakness in front of the mentors, they are external experts who will provide valuable comments from all aspects. Take it as your trial thesis defense...
3. Conference talk. Present ideas and keep the audience awake...

---

Snehal Thakkar (USC) -- Planning for Geospatial Data Integration Motivation: Integrate spatial-related data on the Web, and support query/answering using planning. Subject: Information integration, Geospatial information system Issues: (i) a hierarchical ontology/taxonomy for modeling the spatial application domain, (ii) Planning spatial information query using filtering (reduce querying many irrelevant sources).

Ozgur Simsek (UMass) -- Towards Competence in Autonomous Agents Motivation: Define “useful skills”, and let agents learn them. Subjects: Learning, knowledge discovery Issues: "useful skills" is defined in three categories: (1) access skills, identify "access states" which are critical for making the search space fully searchable especially for hard-to-access regions; (2) approach-discovery skills, how to achieve an "access state"; and (3) causal discovery skill, identify casual relations. Comments: How to memorize the past states is still hard when the problem space scales. Experiences can be reused as well as the skill can be reused.

Mykel J. Kochenderfer (U Edinburgh) -- Adaptive Modeling and Planning for Reactive Agents Motivation: Efficient planning in real time for complex problems with large state and action spaces requires partitioning these spaces into a manageable number of regions. Subjects: reinforcement learning, clustering Issues: Learn to partition the state and action spaces using online split and merge operations. Comments: This could be viewed as a incremental clustering problem such that nodes in trajectories are sample points for generating clusters and thus induce partition of state and action space.

Vincent Conitzer (CMU) -- Computational Aspects of Mechanism Design Motivation: reach optimal outcome by aggregating personal preferences. Subjects: Information aggregation, game theory, multi-item optimization Issues: To achieve optimal outcome for multiple preferences, we can use automated aggregation mechanisms (e.g. vote and auction) and bound agents' behavior. VCG auction encourages users showing true preferences (prevent lying).