Thuật ngữ AI
Từ điển đầy đủ về Trí tuệ nhân tạo
Imagination Learning
Reinforcement learning technique where the agent uses internal models to mentally simulate scenarios and generate experiences without real interaction with the environment. This approach accelerates learning by virtually exploring action trajectories before actual execution.
Mental Planning
Cognitive process where the agent mentally explores and evaluates different action sequences before choosing the best one to execute. This technique uses internal models to anticipate future consequences without real interaction costs.
Environmental Simulation
Artificial generation of environmental experiences using learned models to create additional training data. This method massively increases the data available for the agent's training.
AI Dreaming
Process where an AI agent generates and processes imaginary state sequences during inactive periods to consolidate its learning. These simulated dreams improve model robustness without environmental interaction.
Imagination Experience
Data generated by the agent through mental simulations used as a complement to real experiences for training. These synthetic experiences follow the same structure as real interactions but are produced by the internal model.
Imagined Trajectory
Sequence of simulated states-actions-rewards generated by the agent using its internal environment model. These virtual trajectories allow exploration of alternative policies without real risk or cost.
Imagination Buffer
Memory space storing experiences generated through imagination for the agent's future training. This buffer works in parallel with the real experience buffer to enrich the training corpus.
Intrinsic Curiosity
Internal motivation mechanism that drives the agent to explore unknown or unpredictable states in its environment model. This curiosity guides imagination toward informative scenarios to improve the model.
Model-Guided Exploration
Exploration strategy using the internal model's predictions to identify the most promising actions to try. The agent prioritizes simulations leading to states with high uncertainty or high reward potential.
Virtual Trial-and-Error Learning
Process of improving the policy where the agent tests actions in simulations to learn from their consequences without real interaction. This method eliminates the costs and risks associated with physical exploration.
Future State Prediction
Ability of the internal model to anticipate future states of the environment over multiple time steps. This multi-step prediction is essential for effective imaginative planning.
Environmental Dynamics
Mathematical modeling of the rules governing state transitions in the learning environment. An accurate understanding of this dynamics is crucial for realistic imaginative simulations.
Imaginary Rollout
Simulation of a complete trajectory from a given state using only the internal model and a candidate policy. Imaginary rollouts allow for the rapid evaluation of different action strategies.