Research

The core problem that drives my research is to develop computational agents that exhibit human-level intelligence: real-time systems that persist for long periods of time while autonomously contending with, and improving performance on, a variety of complex problems. General intelligence is a fascinating pursuit in its own right, but also supports diverse applications that are rich, interactive, and adaptive, such as training systems and personal robotic assistants.

Knowledge-Enriched Machine Learning. Bridging dichotomies.

An important capability that humans have is to integrate higher-order knowledge with lower-level cognitive processing in order to make intelligent decisions. Research in this space is challenging and rare, as it requires bridging numerous processing, representational, and political dichotomies that include symbolic vs. sub-symbolic, discrete vs. continuous, and online/incremental vs. batch.

 

Scalable Online Machine Learning

Online machine learning is crucial for many low-level cognitive processes, including real-time perceptual pattern recognition and continuous control. However, few methods can scale to large numbers of examples, in high dimensions, and can support incremental approximation of a wide variety of functions. The Boundary Forest (BF) is a novel instance-based algorithm that satisfies these properties, and, as a bonus, is transparent and easy to implement.

 

Representative publications:

Scalable Optimization via the Three-Weight Message-Passing Algorithm

The Alternating Direction Method of Multipliers (ADMM) is a classic algorithm for convex optimization that has seen an upsurge in recent interest, primarily because it is well suited for distributed implementations. Optimization is a useful framework in the context of cognitive systems because many processes (e.g. constraint satisfaction, planning, vision/perception) can be formulated as optimizing an objective function. While ADMM was developed to solve convex problems, it is well-formed for general optimization, and thus I have been investigating its efficiency and efficacy on a selection of non-convex problems.

 

Representative publications:

Reasoning-Driven Value-Function Learning

Reinforcement Learning (RL) is a key component of many agent architectures, as it is a fully general, online and incremental algorithm that holds many similarities to neural processes in the brain. An RL agent exploits experience in the world to tune a value function, which is a mapping from state-action pairs to an expectation of future reward. The value function then informs action selection, such that the agent maximizes its expected reward. Although RL has been studied extensively, there has been little research on its integration within a cognitive system, nor how the value function is determined by an agent's experience with a task using background knowledge.

 

Representative publications:

Mobile AI. Intelligence in the real world.

Human-level intelligence must contend with the infinite complexities of the real world, including sensing multiple modalities (e.g., vision, audition), localization, and actuation such as to achieve goals.

 

Music Expression and Interaction

How can contemporary work in machine learning and cognitive architecture be used in mobile music interactions? I have integrated the Soar cognitive architecture with the urMus meta-environment and explored learning for music generation and autonomous accompaniment.

 

Representative publications:

Robotic Navigation and Interaction

I have been developing computational and research methods for developing mobile robots that can persist for long periods of time in human-centric environments (e.g., offices, homes).

 

Representative publications:

Long-Term Memory. Effective and efficient access to experience.

A vital aspect of human-level intelligence long-term memory: people are able to accumulate large amounts of experience and access it later in flexible ways, all while dealing with, in real time, the myriad challenges of everyday life. Long-term memory is crucial to human-level reasoning, as illustrated by the deficits in patients afflicted with anterograde amnesia (e.g. H.M.); however, no computational systems support this capability over long periods of time. The work described below has been implemented within the Soar Cognitive Architecture, which is open-source, cross-platform, and comes with comprehensive documentation and tutorials.

 

Semantic Memory

In psychological literature, semantic knowledge is characterized as general information about the world that is independent of the context in which it was originally learned. An agent endowed with a long-term semantic memory is able to access large stores of information (e.g. lexicon, ontology, fact base) that may be useful across many situations.

 

Representative publications:

Episodic Memory

In psychological literature, episodic memory is a temporally organized sequence of detailed events as originally experienced by an agent. Prior work has shown that agents endowed with flexible access to prior experience are enhanced with numerous cognitive capabilities, such as virtual sensing and action modeling. However, prior work has not produced a task-independent episodic memory that can scale real-time performance across long lifetimes.

 

Representative publications:

Forgetting

In the AI literature, the utility problem refers to the situation in which learning more knowledge can harm an agent's problem-solving performance; one common strategy to address this issue is to incrementally forget a subset of learned knowledge. However, prior work has demonstrated the challenge inherent in devising forgetting policies that work well across problem domains, effectively balancing the task performance of agents with reductions in retrieval time and storage requirements of learned knowledge.

 

Representative publications:

Misc. Reports, presentations, tutorials.

Visiting Researcher @ University of Hertfordshire

  •   
    Final Report
    Evaluating Methods for Long-Term HRI Studies Under Resource Constraints

Grad School

  •      
    Dissertation
    Effective and Efficient Memory for Generally Intelligent Agents
  •   
    Proposal
    Long-Term Declarative Memory for Generally Intelligent Agents
  •   
    Prelim
    Efficiently Implementing Episodic Memory in Soar

Presentations

2019
  • Design-ReSOARpe: A Challenge Problem
    The 39th Soar Workshop. Ann Arbor, MI, USA.
2018
  • Enabling Dynamic Agent-Defined Learning Problems in Soar
    The 38th Soar Workshop. Ann Arbor, MI, USA.
2015
  • Mining Episodic Memory
    The 35th Soar Workshop. Ann Arbor, MI, USA.
2014
  • Discussing the Future of Episodic Memory in Soar
    The 34th Soar Workshop. Ann Arbor, MI, USA.
  • The Boundary Forest Algorithm for Fast Online Learning of High-Dimensional Data
    The 34th Soar Workshop. Ann Arbor, MI, USA
  • The Boundary Forest Algorithm for Fast Online Supervised Learning
    The 3rd New England Machine Learning Day (NEML). Cambridge, MA, USA.
2013
  •   
    The Three-Weight Algorithm: Enhancing ADMM for Large-Scale Distributed Optimization
    Advances in Neural Information Processing Systems 26 (NIPS). Lake Tahoe, NV, USA.
  • The Three-Weight Algorithm: A Flexible Platform for Integrating Knowledge and Optimization
    The 33rd Soar Workshop. Ann Arbor, MI, USA.
  • Improved Message-Passing Algorithm Incorporating Certainty Information
    The 2nd New England Machine Learning Day (NEML). Cambridge, MA, USA.
  • Effective and Efficient Memory for Generally Intelligent Agents
    Interactive Robotics Group, MIT. Cambridge, MA.
2012
  • Competence-Preserving Retention of Learned Knowledge in Soar's Working and Procedural Memories
    The 32nd Soar Workshop. Ann Arbor, MI, USA.
  • A Multi-Domain Evaluation of Scaling in Soar's Episodic Memory
    The 32nd Soar Workshop. Ann Arbor, MI, USA.
  • Future Memory Research in Soar
    The 32nd Soar Workshop. Ann Arbor, MI, USA.
  • Effective and Efficient Memory for Generally Intelligent Agents
    Artificial Life, Intelligence and Adaptive Systems (ALAIS) Meeting, University of Hertfordshire. Hatfield, United Kingdom.
2011
  •   
    Effective Scaling of Long-term Memory for Reactive Rule-based AgentsInvited Talk
    The 2011 International Conference on Reasoning Technologies (Rules Fest). San Francisco, CA, USA.
  • SoarQnA: Standardized Access to External Knowledge
    The 31st Soar Workshop. Ann Arbor, MI, USA.
  • The State of Soar: v9.3.1
    The 31st Soar Workshop. Ann Arbor, MI, USA.
  • Soaring to New Platforms: 2011 Update
    The 31st Soar Workshop. Ann Arbor, MI, USA.
  • Efficient Activation-based Working Memory Forgetting
    The 31st Soar Workshop. Ann Arbor, MI, USA.
  • Effective and Efficient Historical Memory Retrieval Bias in Soar's Semantic Memory
    The 31st Soar Workshop. Ann Arbor, MI, USA.
  • Episodic & Semantic: Soar's Long-Term Declarative Memory Systems
    The 31st Soar Workshop. Ann Arbor, MI, USA.
  • Databases & Your Research
    Soar Group, University of Michigan. Ann Arbor, MI, USA.
2010
  • Soaring to New Platforms
    The 30th Soar Workshop. Ann Arbor, MI, USA.
  • Soar2Soar
    The 30th Soar Workshop. Ann Arbor, MI, USA.
  • Large Semantic Stores in Soar
    The 30th Soar Workshop. Ann Arbor, MI, USA.
  • Long-Term Symbolic Memories for Long-Living Learning Agents
    The 30th Soar Workshop. Ann Arbor, MI, USA.
  • Speedy: A Lightweight Platform for Logging, Analyzing, and Visualizing Experimental Data
    Database Research Group, University of Michigan. Ann Arbor, MI, USA.
2009
  • What Does Cyc Know?
    Soar Technology, Inc. Ann Arbor, MI, USA.
  • iSoar: Soar on the iPhone OS
    The 29th Soar Workshop. Ann Arbor, MI, USA.
  • Soar-SMem: A Public Pilot
    The 29th Soar Workshop. Ann Arbor, MI, USA.
  • Efficiently Implementing Episodic Memory
    The 29th Soar Workshop. Ann Arbor, MI, USA.
  • Only Process Changes: Efficiently Implementing Episodic Memory
    Soar Group, University of Michigan. Ann Arbor, MI, USA.
2008
  • Episodic Memory: A DBMS Perspective
    Database Research Group, University of Michigan. Ann Arbor, MI, USA.
  • Soar-RL: A Year of "Learning"
    The 28th Soar Workshop. Ann Arbor, MI, USA.
  • Episodic Memory and Databases: A Year to "Remember"
    The 28th Soar Workshop. Ann Arbor, MI, USA.
  • SoarSim: An Experimentation Framework
    The 28th Soar Workshop. Ann Arbor, MI, USA.

Tutorials

The 39th Soar Workshop. Ann Arbor, MI, USA. (2019)
  • Semantic Memory
  • Episodic Memory
  • Soar Markup Language (SML)
The 38th Soar Workshop. Ann Arbor, MI, USA. (2018)
  • Reinforcement Learning
  • Semantic Memory
  • Episodic Memory
  • Soar Markup Language (SML)
The 25th International Joint Conference on Artificial Intelligence (IJCAI). New York City, NY, USA. (2016)
  • The Soar Cognitive Architecture
The 36th Soar Workshop. Ann Arbor, MI, USA. (2016)
  • Reinforcement Learning
  • Semantic Memory
  • Episodic Memory
  • Soar Markup Language (SML)
The 35th Soar Workshop. Ann Arbor, MI, USA. (2015)
  • Episodic Memory
  • Reinforcement Learning
The 12th International Conference on Cognitive Modeling (ICCM). Ottawa, Canada. (2013)
  • Soar's Spatial Visual System (SVS)
  • Episodic Memory
  • Semantic Memory
  • Reinforcement Learning
The 33rd Soar Workshop. Ann Arbor, MI, USA. (2013)
  • Soar Markup Language (SML)
  • Episodic Memory
  • Semantic Memory
  • Reinforcement Learning
The 32nd Soar Workshop. Ann Arbor, MI, USA. (2012)
  • Soar Markup Language (SML)
  • Episodic Memory
  • Semantic Memory
  • Reinforcement Learning
The 31st Soar Workshop. Ann Arbor, MI, USA. (2011)
  • Soar Markup Language (SML)
  • Episodic Memory
  • Semantic Memory
  • Reinforcement Learning
The 30th Soar Workshop. Ann Arbor, MI, USA. (2010)
  • Episodic Memory
  • Semantic Memory
  • Reinforcement Learning
The 29th Soar Workshop. Ann Arbor, MI, USA. (2009)
  • Episodic Memory
  • Reinforcement Learning