Dynamic Systems as a Multi-Agent Simulation – The recent advances in AI applications have proven to be highly successful. In this paper, we present a system that uses a human-generated video from a mobile phone to perform complex tasks such as action recognition and vision in a robotic arm, as a semi-supervised process. We train the robot to perform multiple, sequential, action-based tasks, based on the action set that human players perform on the video. These tasks are presented as a new feature from the video, which could be used as a proxy to measure cognitive activity. The video captured by the robot shows human players performing multiple actions and actions on different video frames, in order to assess the visual state of the agent. We show how in this way the robotic arm and our video can be integrated to a single, sequential action detection system. In particular, we show how to train an action-tracking system that aims to recognize the actions of each player as a sequence of action clusters. We analyze the results of both the robot and human tasks to demonstrate the effectiveness of the system.

Many computer vision tasks require representing objects as a sequence of sequences, or sequences of objects as a pair of sequences. This work presents a framework for model learning based on dataflow learning to model the behavior of objects of interest (objects that appear within a given target scene). Our framework extends previous work on object-behavior modeling to the multi-task sequential learning problem, which has been studied extensively. By using a recurrent neural network (RNN), our multi-task learning framework allows us to model the behavior of objects, their state spaces, and their behaviors over long time horizons. We show by experiments that our approach outperforms existing methods for multiple-task reinforcement learning or learning sequential models of object behavior in both real and synthetic data.

A Semantics-Driven Approach to Evaluation of Sports Teams’ Ratings from Draft Descriptions

An Online Convex Optimization Approach for Multi-Relational Time Series Prediction

Dynamic Systems as a Multi-Agent Simulation

On the Connectivity between Reinforcement Learning and the Game of Go

Boosting Feature Selection in Multi-Task Multi-Task LearningMany computer vision tasks require representing objects as a sequence of sequences, or sequences of objects as a pair of sequences. This work presents a framework for model learning based on dataflow learning to model the behavior of objects of interest (objects that appear within a given target scene). Our framework extends previous work on object-behavior modeling to the multi-task sequential learning problem, which has been studied extensively. By using a recurrent neural network (RNN), our multi-task learning framework allows us to model the behavior of objects, their state spaces, and their behaviors over long time horizons. We show by experiments that our approach outperforms existing methods for multiple-task reinforcement learning or learning sequential models of object behavior in both real and synthetic data.