Glossary

Action: Each input from your industrial environment that will be controlled by your Autonomous AI system. For example, the three actions in car-racing environment would be: the steering wheel, the acceleration pedal, and the brake.

Algorithm: A process or set of instructions used to perform a task.

Artificial Intelligence: Computer systems that can perform tasks that normally require human intelligence.

Assessment: A set of episodes where the Autonomous AI policy is evaluated on the scenarios of interest where the solution must be effective. Autonomous in the Bonsai platform come in two types: automatic assessments, which are run throughout training to ensure only better policies supercede the existing champion; and custom assessments, which the user runs manually to evaluate the performance and query individual episode data.

Automation: Intelligence applied to industrial tasks, usually using mathematical calculations, optimization, and rule systems

Autonomous artificial intelligence: Any system that perceives its environment and takes actions that maximize its chance of achieving its goals.

Benchmark: A number that represents the effectiveness of current methods. This performance of the current method or baseline policy is measured using the KPI agreed upon with SMEs. For example, a benchmark for a car-racing environment based on expert drivers could be 260 minutes for 20-lap races.

Brain weights: Each of the parameters that define the current Bonsai brain policies. Bonsai training starts with randomized weights. During the training session, these weights are modified towards exploring more effective policies. Once a better policy is found during an Automatic Assessment, the new weights overwrite the previously stored, becoming the new champion.

Challenger: A policy that is being compared against the current champion or best policy found. During brain training, the brain explores several policies. Recurrently, the new policies are assessed through an automatic assessment. When the time comes for an assessment, the current explored policy being assessed becomes the challenger.

Champion: The best policy found over an ongoing or finished training session. Champions are initialized with a random policy, and are only overwritten when a new policy (a challenger) is found to perform better.

Closed-loop: A system that can both take action and receive feedback.

Concept: A skill that the brain can learn independently. Concepts are the components that permit building modular brains. The decomposition of a monolithic brain into several concepts can happen based on strategy (such as learning to take turns versus accelerating and passing on straight lanes); or based on physical independence (such as learning to race with a convertible versus an SUV).

Control Theory: A method of automated decision making that uses math to achieve predictable results.

Data Driven Simulation: A simulation that uses machine learning or other data driven methods to predict what will happen in the system when a specific action is taken.

Deep Learning: A machine learning method where learning is stored in a deep neural network.

Deep Reinforcement Learning: A machine learning method where AI learns through practice and the results of learning are stored in a deep neural network.

Episode: A single scenario in Autonomous AI training. For example, one episode of a car race would be a lap and one episode of a game would be a turn, one episode in manufacturing might be a “run” that produces a specific product for a period of time.

Expert Policy: The control strategies that operators follow today based on their current method of operation. An autonomous AI solution can be considered an expert policy once it has been vetted to outperform the current method’s performance without running into unsafe/undesired conditions.

Expert Rules: Directions for taking action based on previous experience.

Expert Systems: Intelligent systems that rely on a database of expert rules, a user interface, and an inference engine for interpreting the rules.

Feedback Control: Automated decision-making that calculates the next decision based on feedback.

Feedforward Control: Automated decision-making that controls for disturbances and for the control variable separately.

First Principles (simulation): A simulation that models the real system with math and science equations.

Goal Objective: Each of the individual conditions that must be met during goal definition. These objectives are evaluated individually during an episode for each applicable scenarios.

Goal Robustness: A continuous and unbounded metric to measure effectiveness on a goal-based concept or brain. This metric is continuous, enabling users to track even the slightest improvements on performance. This metric is also unbounded, which means that it will show performance improvements beyond the defined success thresholds for each individual goal objective.

Goal Satisfaction: A continuous metric to measure effectiveness on a goal-based concept or brain. This metric is continuous, enabling users to track even the slightest improvements on performance. This metric is bounded by a value of 100%, which is reached when the champion meets the success thresholds defined across all individual goal objectives.

Goal: A high-level specification of what you want the AI to learn. Each goal identifies all the objectives that the concept needs to learn.

Inkling: A domain specific programming language for machine teaching.

Iteration: One control action or decision. Iterations string together into episodes.

Key performance indicator (KPI): A metric used to evaluate success for the scenarios of interest. This metric must be agreed upon with SMEs and cannot entail any subjective assertions. The metric or formula is usually in the units that the business and SMEs care about. For example, in a car-racing environment, the KPI could be the total time taken to run a 20-lap race.

Learning Layer: The underlying AI algorithms that are learning during brain training sessions. The algorithms of the learning layer are the code and content that are exported for deployment. (See also, Teaching Layer)

Machine Learning: Algorithms that improve automatically through experience and by seeing data.

Machine Teaching: A framework for using human teaching methods to help machines learn concepts and acquire skills.

Module: A “drag and drop” component of the brain. Modules can be ‘Learned’, ‘Selector’, ‘Programmed’, and ‘Imported’. A module contains a concept or skill and the requisite goals the brain needs to learn.

MPC Control: Automated decision-making that tries potential actions in advance using optimization algorithms and an accurate system model.

Open-loop: Automated decision-making based on a prescribed mathematical relationship.

Optimization: Algorithms that search for potential solutions.

Over-fitting: A condition where a champion policy is over-performing in one particular scenario, while under-performing in other scenarios.

PID Control: Feedback control that uses three gains to determine the controller behavior.

Policy (over-performing): Policies that beat the benchmark and/or meet our success criteria. Sometimes, also used to refer to champions, which have better performance over any past champion.

Policy (see also: expert policy): Any control strategy Autonomous or otherwise that takes the environment states as inputs, and outputs the corresponding values for each environment variable. These controls do not need to be effective. A new brain, randomly initialized, constitutes a valid policy for the environment. A policy defined by current methods (including operators or not) is called an Expert Policy.

Policy (under-performing): Policies that do not beat the benchmark and/or do not meet our success criteria.

Reinforcement Learning: A machine learning method where AI practices performing a task.

Scenario: A set of environmental state conditions.

Simulation: A digital or virtual version of a real system. For Autonomous AI, this provides the training gym for the AI to practice taking action and receiving feedback.

Skill: Skills are units of competence (ability to achieve goals) at a specific task. They are the building blocks of Autonomous AI.

Strategy: A labeled course of action for completing a task.

Struct: A jargon term for structure referring to sections of Inkling code within Bonsai.

Subject Matter Experts (SMEs): Humans that understand aspects of the problem being solved. Some SMEs might be operators involved in the process, while other SMEs might be experts that build the simulations that train the brain to solve the problem.

Surrogate Model (simulation): A simplified version of a simulation that can run faster than a full simulation. Surrogate models simulations are used in early brain training to ‘coarse’ tune a brain; while full simulations are used in later brain training to ‘fine’ tune the brain.

Teaching Layer: The totality of the Bonsai interface, including the organization of the brain’s modules, and the goals, lessons, and parameters used to influence the performance of the learning layer. (See also: Learning Layer)