Reinforcement learning with adaptive structure
Reinforcement learning with adaptive structure (RLAS) describes learning from decision outcomes while revising the structure used to evaluate choices. Measured outcomes teach action preferences and value estimates. Eligible experience also shapes the representations, rules, model parameters, and predictor relationships that contribute to later decisions. Each event updates only the mechanisms eligible for its evidence and task contract. Action credit connects an outcome to the choice that produced it. Sequential learning lets later rewards revise earlier state-changing actions. Structural credit concerns which prediction sources and combinations contributed useful information. CUP learns that contribution when enabled. Two sources can each be useful while their combination is harmful; supported utility can preserve individual contributions and inhibit the combination. Plasticity rules explains this distinction. The resulting state persists across interactions. Later decisions use the revised representations, models, rules, and action values. During frozen use, current observations are processed through retained state with learner updates disabled.Form useful task structure
With Discovery, inputs, field combinations, temporal context, and action or intervention bindings can begin unresolved. Adapt-1 forms an executable learner view from interaction and retains it as revisable state. Compatible earlier events can then be reconsidered through the formed view. Structural learning induces rules from observations. Later evidence can specialize, generalize, replace, or retire them. Recurring combinations of active rules can form recognizable latent states, allowing their learned transition history to be reused when the same rule configuration returns. Discovery covers structure formation.Reason with retained evidence
Relational Hopfield Memory combines similarity with associative recovery. Typed relations and hyperedges preserve the joint context of observations. Executable conditions and hypotheses support comparison of explanations and inference of consequences. Hypothesis results expose their supporting evidence and counterevidence. Adaptive reads can strengthen associations. Consolidation stabilizes selected records, and corrections revise their evidence. Unused associations lose influence. Later queries use this evolving state to recover context and produce supported results. See Continuity across calls and Explain a result.Learn predictions and decisions
Transition learning estimates observable consequences. Contextual policies learn which choices work in a given situation, and sequential credit allows later outcomes to revise earlier state-changing choices. Configured transition planning evaluates candidate action sequences against a goal and constraints using learned predictions of successive states. Completed executions supply new transition evidence for later plans. Native planning compares continuations from exact rules, learned forecasts, or a simulator. A transition model predicts action effects; a value model estimates the utility of unfinished states. Exact terminal scores can come from the task rules or objective. Learning and planning describes how these values enter native search and how the application executes its choice. When change detection and historical-regime matching are enabled, Core can recognize shifts in operating conditions and reuse retained regimes that match new evidence. Configured support requirements determine when a historical match can be used. Adapt-1 updates the applicable state as observations and attributed outcomes arrive. Where trainable models are enabled, Core trains candidates from accumulated evidence, validates them on later observations, and installs accepted models while operation continues. The Preview article describes neural evidence rerankers, link predictors, and nonlinear policy models trained within Core. Accepted model parameters become part of retained learned state. Shared parameters mean that training on one input can change predictions for other inputs. The learning and planning article describes value-parameter fitting, validation, adoption, and continued learning within the substrate. Responses can identify participating learners and model revisions. What a trace exposes explains how these diagnostics relate to the internal learned state. See Learning relationships for the supported evidence and feedback patterns, or Learn from an outcome to connect execution to later decisions.Acquire and revise procedures
In Machina, Adapt-1 operates with continuous control sequences. It acquires commands through execution and measured outcomes, compares attempts, tests command removal and ordering, and learns contextual corrections to an acquired base. The resulting procedure becomes reusable learned state. During frozen use, current observations can yield different controls through unchanged learned parameters. See Machina and Retained use. Learning schedules explains how to configure acquisition and later use, including runs that keep the learner state fixed.Mechanism maturity
Numeric sequence acquisition is currently markedexperimental: true by the trajectory API. The acquisition guide describes the supported proposal, execution, and observation workflow. Check the mechanism status and request-contract coverage for the deployment your application uses.
Optional mechanisms
Optional mechanisms
Temporal Context Projection (TCP) makes bounded episode history available when the current observation alone is insufficient.Counterfactual Utility Plasticity (CUP) learns which registered prediction sources and combinations add useful predictive information, then revises their influence on later predictions as evidence changes.