AI Glossary
The complete dictionary of Artificial Intelligence
Differentiable Transition Model
Mathematical function describing the evolution of the state of a continuous system, designed to be differentiable to allow optimization through gradient descent in reinforcement learning algorithms.
Stochastic Ordinary Differential Equations (ODE)
System of differential equations incorporating a random noise term, used to model the uncertain dynamics of continuous environments while maintaining the differentiability necessary for learning.
Differentiable Numerical Integrator
Numerical computation method (e.g., Euler, Runge-Kutta) whose implementation is differentiable, allowing gradients to be propagated through time simulation steps for the optimization of dynamic models.
Radial Basis Function (RBF) Neural Network
Neural network architecture using radial basis functions as activation functions, particularly suitable for approximating continuous and differentiable functions for dynamics modeling.
Optimized Trajectory Planning (TPO)
Planning method in the trajectory space that directly optimizes an action sequence using a differentiable model, with updates based on expected reward gradients.
Hamiltonian Systems Modeling
Continuous dynamics modeling approach based on the energy conservation principles of Hamiltonian systems, ensuring long-term stability and differentiability properties.
Automatic Differentiation through Time
Gradient computation technique that propagates backpropagation through the time steps of a continuous simulation, essential for training differentiable dynamics models.
Temporal Gaussian Process Model (TGPM)
Extension of Gaussian processes for continuous time series modeling, providing calibrated uncertainty while maintaining differentiability for optimization in reinforcement learning.
Differentiable Neural Controller
Neural network implementing a control policy whose outputs are differentiable functions of the input states, enabling joint optimization with the dynamics model in model-based frameworks.
Differentiable Multiple Shooting Method
Algorithm for solving boundary value problems for continuous systems, adapted to be differentiable and thus allowing parameter optimization in reinforcement learning trajectories.
Basis Function State Space Model
Representation of continuous dynamics where state transitions are approximated by a linear combination of differentiable basis functions, facilitating analytical optimization of model parameters.
Differentiable Model Policy Optimization (DMPO)
Variant of policy optimization where gradients are computed through a differentiable dynamics model, combining the advantages of model-based and model-free methods for continuous environments.
Learned Dynamics Equation (LDE)
Mathematical formulation where the parameters of a differential equation describing system dynamics are learned through optimization, while preserving the differentiable structure of the original equation.
Differentiable Continuous-Discrete Hybrid Model
Modeling architecture combining differentiable continuous components with discrete events, where transitions are smoothed to maintain overall system differentiability.
Differentiable Integration State Prediction
Process of predicting future states using numerical integration where the operation itself is differentiable, allowing computation of gradients of the prediction with respect to model parameters.
Physics-Informed Neural Network (PINN)
Neural architecture that incorporates differential equations from physics into its loss function, ensuring the learned model respects conservation laws while remaining differentiable.
Differentiable Collocation Method
A technique for solving constrained optimization problems for continuous systems, where collocation constraints are formulated as differentiable functions for policy training.
Navier-Stokes Equation Transition Model
Use of Navier-Stokes equations, made differentiable through appropriate discretization, to model fluid dynamics in continuous reinforcement learning environments.
Differentiable Augmented Lagrangian Optimization
A constrained optimization method where the augmented Lagrangian function is differentiable with respect to state and control variables, enabling its use in reinforcement learning loops.