Simulating Behaviors of POMDPs
Advisor(s)
Piotr Gmytrasiewicz, UIC Artificial Intelligence Laboratory
Location
Room B1018-2
Start Date
26-4-2019 11:05 AM
End Date
26-4-2019 11:20 AM
Abstract
Intuition is an important human capability that allows us to gauge and predict the properties of an unknown environment. This study presents an approach to how a computer agent can replicate this innate intuition by using an algorithm that will learn from an unknown underlying state using Partially Observable Markov Decision Processes (POMDPs). This task is difficult because the computer will never have 100% certainty in its decisions, thus every action will be based on a belief state. To determine the transition probabilities in our unknown environment, we are currently recursively applying a learning algorithm, looking for patterns in the computer’s belief state. We are implementing these ideas and algorithms using Java in an effort to create an intelligent machine. We test the competence of our algorithms by running them in a simulation with known transition probabilities and finding percent error between calculated probabilities and true probabilities. Although a robust, error-free algorithm has not been developed, we hope to find a logical and consistent algorithm that will determine transition and observational probabilities accurately and consistently.
Simulating Behaviors of POMDPs
Room B1018-2
Intuition is an important human capability that allows us to gauge and predict the properties of an unknown environment. This study presents an approach to how a computer agent can replicate this innate intuition by using an algorithm that will learn from an unknown underlying state using Partially Observable Markov Decision Processes (POMDPs). This task is difficult because the computer will never have 100% certainty in its decisions, thus every action will be based on a belief state. To determine the transition probabilities in our unknown environment, we are currently recursively applying a learning algorithm, looking for patterns in the computer’s belief state. We are implementing these ideas and algorithms using Java in an effort to create an intelligent machine. We test the competence of our algorithms by running them in a simulation with known transition probabilities and finding percent error between calculated probabilities and true probabilities. Although a robust, error-free algorithm has not been developed, we hope to find a logical and consistent algorithm that will determine transition and observational probabilities accurately and consistently.