Filtering by
- All Subjects: Machine Learning
- All Subjects: Electrical Engineering
The research presented in this Honors Thesis provides development in machine learning models which predict future states of a system with unknown dynamics, based on observations of the system. Two case studies are presented for (1) a non-conservative pendulum and (2) a differential game dictating a two-car uncontrolled intersection scenario. In the paper we investigate how learning architectures can be manipulated for problem specific geometry. The result of this research provides that these problem specific models are valuable for accurate learning and predicting the dynamics of physics systems.<br/><br/>In order to properly model the physics of a real pendulum, modifications were made to a prior architecture which was sufficient in modeling an ideal pendulum. The necessary modifications to the previous network [13] were problem specific and not transferrable to all other non-conservative physics scenarios. The modified architecture successfully models real pendulum dynamics. This case study provides a basis for future research in augmenting the symplectic gradient of a Hamiltonian energy function to provide a generalized, non-conservative physics model.<br/><br/>A problem specific architecture was also utilized to create an accurate model for the two-car intersection case. The Costate Network proved to be an improvement from the previously used Value Network [17]. Note that this comparison is applied lightly due to slight implementation differences. The development of the Costate Network provides a basis for using characteristics to decompose functions and create a simplified learning problem.<br/><br/>This paper is successful in creating new opportunities to develop physics models, in which the sample cases should be used as a guide for modeling other real and pseudo physics. Although the focused models in this paper are not generalizable, it is important to note that these cases provide direction for future research.
High-entropy alloys possessing mechanical, chemical, and electrical properties that far exceed those of conventional alloys have the potential to make a significant impact on many areas of engineering. Identifying element combinations and configurations to form these alloys, however, is a difficult, time-consuming, computationally intensive task. Machine learning has revolutionized many different fields due to its ability to generalize well to different problems and produce computationally efficient, accurate predictions regarding the system of interest. In this thesis, we demonstrate the effectiveness of machine learning models applied to toy cases representative of simplified physics that are relevant to high-entropy alloy simulation. We show these models are effective at learning nonlinear dynamics for single and multi-particle cases and that more work is needed to accurately represent complex cases in which the system dynamics are chaotic. This thesis serves as a demonstration of the potential benefits of machine learning applied to high-entropy alloy simulations to generate fast, accurate predictions of nonlinear dynamics.
Intelligent agents learn from experiences, and in times of uncertainty use the knowl-
edge acquired to make decisions and accomplish their individual or team objectives.
Agent objectives are defined using cost functions designed uniquely for the collective
task being performed. Individual agent costs are coupled in such a way that group ob-
jective is attained while minimizing individual costs. Information Asymmetry refers
to situations where interacting agents have no knowledge or partial knowledge of cost
functions of other agents. By virtue of their intelligence, i.e., by learning from past
experiences agents learn cost functions of other agents, predict their responses and
act adaptively to accomplish the team’s goal.
Algorithms that agents use for learning others’ cost functions are called Learn-
ing Algorithms, and algorithms agents use for computing actuation (control) which
drives them towards their goal and minimize their cost functions are called Control
Algorithms. Typically knowledge acquired using learning algorithms is used in con-
trol algorithms for computing control signals. Learning and control algorithms are
designed in such a way that the multi-agent system as a whole remains stable during
learning and later at an equilibrium. An equilibrium is defined as the event/point
where cost functions of all agents are optimized simultaneously. Cost functions are
designed so that the equilibrium coincides with the goal state multi-agent system as
a whole is trying to reach.
In collective load transport, two or more agents (robots) carry a load from point
A to point B in space. Robots could have different control preferences, for example,
different actuation abilities, however, are still required to coordinate and perform
load transport. Control preferences for each robot are characterized using a scalar
parameter θ i unique to the robot being considered and unknown to other robots.
With the aid of state and control input observations, agents learn control preferences
of other agents, optimize individual costs and drive the multi-agent system to a goal
state.
Two learning and Control algorithms are presented. In the first algorithm(LCA-
1), an existing work, each agent optimizes a cost function similar to 1-step receding
horizon optimal control problem for control. LCA-1 uses recursive least squares as
the learning algorithm and guarantees complete learning in two time steps. LCA-1 is
experimentally verified as part of this thesis.
A novel learning and control algorithm (LCA-2) is proposed and verified in sim-
ulations and on hardware. In LCA-2, each agent solves an infinite horizon linear
quadratic regulator (LQR) problem for computing control. LCA-2 uses a learning al-
gorithm similar to line search methods, and guarantees learning convergence to true
values asymptotically.
Simulations and hardware implementation show that the LCA-2 is stable for a
variety of systems. Load transport is demonstrated using both the algorithms. Ex-
periments running algorithm LCA-2 are able to resist disturbances and balance the
assumed load better compared to LCA-1.