This thesis explores how biologically plausible learning algorithms can build internal representations in the brain to enable adaptive behavior. The first part examines reinforcement learning (RL) with multiple temporal discounts, allowing agents to learn the full temporal evolution of reward probabilities rather than aggregated values. This approach aligns with experimental findings on dopamine activity in mice. The second part investigates internal representations for robust motor control using a musculoskeletal model of the human hand in an object manipulation task. A curriculum learning approach is developed, and the resulting control signals are analyzed for similarities with human muscle coordination. The final part addresses how animals achieve rapid learning with minimal experience, proposing a framework where behavioral primitives enable predictive simulations, vastly improving sample efficiency over traditional RL.