Engineering topics

Agogino, Adrian K.

Publications and source records attributed to Agogino, Adrian K..

16 records

NASA NTRS2020

Improving Trust in Deep Neural Networks with Nearest Neighbors

Deep neural networks are used increasingly for perception and decision-making in UAVs. For example, they can be used to recognize objects from images and decide what actions the vehicle should take. While deep neural networks can perform very well at complex tasks, their decisions may be unintuitive to a human operator. When a human disagrees with a neural network prediction, due to the black box nature of deep neural networks, it can be unclear whether the system knows something the human does not or whether the system is malfunctioning. This uncertainty is problematic when it comes to ensuring safety. As a result, it is important to develop technologies for explaining neural network decisions for trust and safety. This paper explores a modification to the deep neural network classification layer to produce both a predicted label and an explanation to support its prediction. Specifically, at test time, we replace the final output layer of the neural network classifier by a k-nearest neighbor classifier. The nearest neighbor classifier produces 1) a predicted label through voting and 2) the nearest neighbors involved in the prediction, which represent the most similar examples from the training dataset. Because prediction and explanation are derived from the same underlying process, this approach guarantees that the explanations are always relevant to the predictions. We demonstrate the approach on a convolutional neural network for a UAV image classification task. We perform experiments using a forest trail image dataset and show empirically that the hybrid classifier can produce intuitive explanations without loss of predictive performance compared to the original neural network. We also show how the approach can be used to help identify potential issues in the network and training process.

Lee, Ritchie↗

NASA NTRS2018

Super Ball Bot - Structures for Planetary Landing and Exploration

Small, light-weight and low-cost missions will become increasingly important to NASA's exploration goals for our solar system. Ideally teams of dozens or even hundreds of small, collapsable robots, weighing only a few kilograms a piece, will be conveniently packed during launch and would reliably separate and unpack at their destination. Such teams will allow rapid, reliable in-situ exploration of hazardous destination such as Titan, where imprecise terrain knowledge and unstable precipitation cycles make single-robot exploration problematic. Unfortunately landing many lightweight conventional robots is difficult with conventional technology. Current robot designs are delicate, requiring combinations of devices such as parachutes, retrorockets and impact balloons to minimize impact forces and to place a robot in a proper orientation. Instead we propose to develop a radically different robot based on a "tensegrity" built purely upon tensile and compression elements. These robots can be light-weight, absorb strong impacts, are redundant against single-point failures, can recover from different landing orientations and are easy to collapse and uncollapse. We believe tensegrity robot technology can play a critical role in future planetary exploration.

Agogino, Adrian K.↗

NASA NTRS2017

Controlling Tensegrity Robots through Evolution using Friction based Actuation

Traditional robotic structures have limitations in planetary exploration as their rigid structural joints are prone to damage in new and rough terrains. In contrast, robots based on tensegrity structures, composed of rods and tensile cables, offer a highly robust, lightweight, and energy efficient solution over traditional robots. In addition tensegrity robots can be highly configurable by rearranging their topology of rods, cables and motors. However, these highly configurable tensegrity robots pose a significant challenge for locomotion due to their complexity. This study investigates a control pattern for successful locomotion in tensegrity robots through an evolutionary algorithm. A twelve-rod hardware model is rapidly prototyped to utilize a new actuation method based on friction. A web-based physics simulation is created to model the twelve-rod tensegrity ball structure. Square-waves are used as control policies for the actuators of the tensegrity structure. Monte Carlo trials are run to find the most successful number of amplitudes for the square-wave control policy. From the results, an evolutionary algorithm is implemented to find the most optimized solution for locomotion of the twelve-rod tensegrity structure. The software pattern coupled with the new friction based actuation method can serve as the basis for highly efficient tensegrity robots in space exploration.

Tensegrit↗

NASA NTRS2014

Hardware Design and Testing of SUPERball, A Modular Tensegrity Robot

We are developing a system of modular, autonomous "tensegrity end-caps" to enable the rapid exploration of untethered tensegrity robot morphologies and functions. By adopting a self-contained modular approach, different end-caps with various capabilities (such as peak torques, or motor speeds), can be easily combined into new tensegrity robots composed of rods, cables, and actuators of different scale (such as in length, mass, peak loads, etc). As a first step in developing this concept, we are in the process of designing and testing the end-caps for SUPERball (Spherical Underactuated Planetary Exploration Robot), a project at the Dynamic Tensegrity Robotics Lab (DTRL) within NASA Ames's Intelligent Robotics Group. This work discusses the evolving design concepts and test results that have gone into the structural, mechanical, and sensing aspects of SUPERball. This representative tensegrity end-cap design supports robust and repeatable untethered mobility tests of the SUPERball, while providing high force, high displacement actuation, with a low-friction, compliant cabling system.

Tensegrity↗

NASA NTRS2014

Design and Control of Modular Spine-Like Tensegrity Structures

We present a methodology enabled by the NASA Tensegrity Robotics Toolkit (NTRT) for the rapid structural design of tensegrity robots in simulation and an approach for developing control systems using central pattern generators, local impedance controllers, and parameter optimization techniques to determine effective locomotion strategies for the robot. Biomimetic tensegrity structures provide advantageous properties to robotic locomotion and manipulation tasks, such as their adaptability and force distribution properties, flexibility, energy efficiency, and access to extreme terrains. While strides have been made in designing insightful static biotensegrity structures, gaining a clear understanding of how a particular structure can efficiently move has been an open problem. The tools in the NTRT enable the rapid exploration of the dynamics of a given morphology, and the links between structure, controllability, and resulting gait efficiency. To highlight the effectiveness of the NTRT at this exploration of morphology and control, we will provide examples from the designs and locomotion of four different modular spine-like tensegrity robots.

Modular↗

NASA NTRS2014

Evolutionary Agent-Based Simulation of the Introduction of New Technologies in Air Traffic Management

Accurate simulation of the effects of integrating new technologies into a complex system is critical to the modernization of our antiquated air traffic system, where there exist many layers of interacting procedures, controls, and automation all designed to cooperate with human operators. Additions of even simple new technologies may result in unexpected emergent behavior due to complex human/ machine interactions. One approach is to create high-fidelity human models coming from the field of human factors that can simulate a rich set of behaviors. However, such models are difficult to produce, especially to show unexpected emergent behavior coming from many human operators interacting simultaneously within a complex system. Instead of engineering complex human models, we directly model the emergent behavior by evolving goal directed agents, representing human users. Using evolution we can predict how the agent representing the human user reacts given his/her goals. In this paradigm, each autonomous agent in a system pursues individual goals, and the behavior of the system emerges from the interactions, foreseen or unforeseen, between the agents/actors. We show that this method reflects the integration of new technologies in a historical case, and apply the same methodology for a possible future technology.

Simulation↗

NASA NTRS2006

QUICR-learning for Multi-Agent Coordination

Coordinating multiple agents that need to perform a sequence of actions to maximize a system level reward requires solving two distinct credit assignment problems. First, credit must be assigned for an action taken at time step t that results in a reward at time step t > t. Second, credit must be assigned for the contribution of agent i to the overall system performance. The first credit assignment problem is typically addressed with temporal difference methods such as Q-learning. The second credit assignment problem is typically addressed by creating custom reward functions. To address both credit assignment problems simultaneously, we propose the "Q Updates with Immediate Counterfactual Rewards-learning" (QUICR-learning) designed to improve both the convergence properties and performance of Q-learning in large multi-agent problems. QUICR-learning is based on previous work on single-time-step counterfactual rewards described by the collectives framework. Results on a traffic congestion problem shows that QUICR-learning is significantly better than a Q-learner using collectives-based (single-time-step counterfactual) rewards. In addition QUICR-learning provides significant gains over conventional and local Q-learning. Additional results on a multi-agent grid-world problem show that the improvements due to QUICR-learning are not domain specific and can provide up to a ten fold increase in performance over existing methods.

Agogino, Adrian K.↗

NASA NTRS2005

Quicker Q-Learning in Multi-Agent Systems

Multi-agent learning in Markov Decisions Problems is challenging because of the presence ot two credit assignment problems: 1) How to credit an action taken at time step t for rewards received at t' greater than t; and 2) How to credit an action taken by agent i considering the system reward is a function of the actions of all the agents. The first credit assignment problem is typically addressed with temporal difference methods such as Q-learning OK TD(lambda) The second credit assi,onment problem is typically addressed either by hand-crafting reward functions that assign proper credit to an agent, or by making certain independence assumptions about an agent's state-space and reward function. To address both credit assignment problems simultaneously, we propose the Q Updates with Immediate Counterfactual Rewards-learning (QUICR-learning) designed to improve both the convergence properties and performance of Q-learning in large multi-agent problems. Instead of assuming that an agent s value function can be made independent of other agents, this method suppresses the impact of other agents using counterfactual rewards. Results on multi-agent grid-world problems over multiple topologies show that QUICR-learning can achieve up to thirty fold improvements in performance over both conventional and local Q-learning in the largest tested systems.

Agogino, Adrian K.↗

NASA NTRS2005

Multi Agent Reward Analysis for Learning in Noisy Domains

In many multi agent learning problems, it is difficult to determine, a priori, the agent reward structure that will lead to good performance. This problem is particularly pronounced in continuous, noisy domains ill-suited to simple table backup schemes commonly used in TD(lambda)/Q-learning. In this paper, we present a new reward evaluation method that allows the tradeoff between coordination among the agents and the difficulty of the learning problem each agent faces to be visualized. This method is independent of the learning algorithm and is only a function of the problem domain and the agents reward structure. We then use this reward efficiency visualization method to determine an effective reward without performing extensive simulations. We test this method in both a static and a dynamic multi-rover learning domain where the agents have continuous state spaces and where their actions are noisy (e.g., the agents movement decisions are not always carried out properly). Our results show that in the more difficult dynamic domain, the reward efficiency visualization method provides a two order of magnitude speedup in selecting a good reward. Most importantly it allows one to quickly create and verify rewards tailored to the observational limitations of the domain.

Tumer, Kagan↗

NASA NTRS2004

Overcoming Communication Restrictions in Collectives

Many large distributed system are characterized by having a large number of components (eg., agents, neurons) whose actions and interactions determine a %orld utility which rates the performance of the overall system. Such collectives are often subject to communication restrictions, making it difficult for components which try to optimize their own private utilities, to take actions that also help optimize the world utility. In this article we address that coordination problem and derive four utility functions which present different compromises between how aligned a component s private utility is with the world utility and how readily that component can determine the actions that optimize its utility. The results show that the utility functions specifically derived to operate under communication restrictions outperform both traditional methods and previous collective-based methods by up to 75%.

Tumer, Kagan↗

NASA NTRS2004

Time-Extended Policies in Mult-Agent Reinforcement Learning

Reinforcement learning methods perform well in many domains where a single agent needs to take a sequence of actions to perform a task. These methods use sequences of single-time-step rewards to create a policy that tries to maximize a time-extended utility, which is a (possibly discounted) sum of these rewards. In this paper we build on our previous work showing how these methods can be extended to a multi-agent environment where each agent creates its own policy that works towards maximizing a time-extended global utility over all agents actions. We show improved methods for creating time-extended utilities for the agents that are both "aligned" with the global utility and "learnable." We then show how to crate single-time-step rewards while avoiding the pi fall of having rewards aligned with the global reward leading to utilities not aligned with the global utility. Finally, we apply these reward functions to the multi-agent Gridworld problem. We explicitly quantify a utility's learnability and alignment, and show that reinforcement learning agents using the prescribed reward functions successfully tradeoff learnability and alignment. As a result they outperform both global (e.g., team games ) and local (e.g., "perfectly learnable" ) reinforcement learning solutions by as much as an order of magnitude.

Tumer, Kagan↗

NASA NTRS2004

Unifying Temporal and Structural Credit Assignment Problems

Single-agent reinforcement learners in time-extended domains and multi-agent systems share a common dilemma known as the credit assignment problem. Multi-agent systems have the structural credit assignment problem of determining the contributions of a particular agent to a common task. Instead, time-extended single-agent systems have the temporal credit assignment problem of determining the contribution of a particular action to the quality of the full sequence of actions. Traditionally these two problems are considered different and are handled in separate ways. In this article we show how these two forms of the credit assignment problem are equivalent. In this unified frame-work, a single-agent Markov decision process can be broken down into a single-time-step multi-agent process. Furthermore we show that Monte-Carlo estimation or Q-learning (depending on whether the values of resulting actions in the episode are known at the time of learning) are equivalent to different agent utility functions in a multi-agent system. This equivalence shows how an often neglected issue in multi-agent systems is equivalent to a well-known deficiency in multi-time-step learning and lays the basis for solving time-extended multi-agent problems, where both credit assignment problems are present.

Agogino, Adrian K.↗

NASA NTRS2004

Team Formation in Partially Observable Multi-Agent Systems

Sets of multi-agent teams often need to maximize a global utility rating the performance of the entire system where a team cannot fully observe other teams agents. Such limited observability hinders team-members trying to pursue their team utilities to take actions that also help maximize the global utility. In this article, we show how team utilities can be used in partially observable systems. Furthermore, we show how team sizes can be manipulated to provide the best compromise between having easy to learn team utilities and having them aligned with the global utility, The results show that optimally sized teams in a partially observable environments outperform one team in a fully observable environment, by up to 30%.

Agogino, Adrian K.↗

NASA NTRS2003

Team Formation and Communication Restrictions in Collectives

A collective of agents often needs to maximize a "world utility" function which rates the performance of an entire system, while subject to communication restrictions among the agents. Such communication restrictions make it difficult for agents which try to pursue their own "private" utilities to take actions that also help optimize the world utility. Team formation presents a solution to this problem, where by joining other agents, an agent can significantly increase its knowledge about the environment and improve its chances of both optimizing its own utility and that its doing so will contribute to the world utility. In this article we show how utilities that have been previously shown to be effective in collectives can be modified to be more effective in domains with moderate communication restrictions resulting in performance improvements of up to 75%. Additionally we show that even severe communication constraints can be overcome by forming teams where each agent of a team shares the same utility, increasing performance an additional 25%. We show that utilities and team sizes can be manipulated to form the best compromise between how "aligned" an agent s utility is with the world utility and how easily an agent can learn that utility.

Agogino, Adrian K.↗

NASA NTRS2002

Time-Extended Payoffs for Collectives of Autonomous Agents

A collective is a set of self-interested agents which try to maximize their own utilities, along with a a well-defined, time-extended world utility function which rates the performance of the entire system. In this paper, we use theory of collectives to design time-extended payoff utilities for agents that are both aligned with the world utility, and are "learnable", i.e., the agents can readily see how their behavior affects their utility. We show that in systems where each agent aims to optimize such payoff functions, coordination arises as a byproduct of the agents selfishly pursuing their own goals. A game theoretic analysis shows that such payoff functions have the net effect of aligning the Nash equilibrium, Pareto optimal solution and world utility optimum, thus eliminating undesirable behavior such as agents working at cross-purposes. We then apply collective-based payoff functions to the token collection in a gridworld problem where agents need to optimize the aggregate value of tokens collected across an episode of finite duration (i.e., an abstracted version of rovers on Mars collecting scientifically interesting rock samples, subject to power limitations). We show that, regardless of the initial token distribution, reinforcement learning agents using collective-based payoff functions significantly outperform both natural extensions of single agent algorithms and global reinforcement learning solutions based on "team games".

Tumer, Kagan↗

NASA NTRS2001

Learning Sequences of Actions in Collectives of Autonomous Agents

In this paper we focus on the problem of designing a collective of autonomous agents that individually learn sequences of actions such that the resultant sequence of joint actions achieves a predetermined global objective. We are particularly interested in instances of this problem where centralized control is either impossible or impractical. For single agent systems in similar domains, machine learning methods (e.g., reinforcement learners) have been successfully used. However, applying such solutions directly to multi-agent systems often proves problematic, as agents may work at cross-purposes, or have difficulty in evaluating their contribution to achievement of the global objective, or both. Accordingly, the crucial design step in multiagent systems centers on determining the private objectives of each agent so that as the agents strive for those objectives, the system reaches a good global solution. In this work we consider a version of this problem involving multiple autonomous agents in a grid world. We use concepts from collective intelligence to design goals for the agents that are 'aligned' with the global goal, and are 'learnable' in that agents can readily see how their behavior affects their utility. We show that reinforcement learning agents using those goals outperform both 'natural' extensions of single agent algorithms and global reinforcement, learning solutions based on 'team games'.

Turner, Kagan↗