<p>Illustration of Markov Decision Process (MDP) computation using backward induction (dynamic programming).</p>

The goal was to stay alive, so transitions to the absorbing zero-energy state (in red) incur a reward of –1, while all other transitions have a reward of 0. Crucially, transitions between states are governed by the probability of gain (p with ) and the risk of threat <b>(r)</b>, which determine how likely actions are to lead to different future energy states. <b>(B)</b> Backward induction illustrated schematically.

Value updates proceed from the last to the first time-point (“days”), iteratively combining expected outcomes and selecting the higher-valued action at each step. State–action values (Q-values) represent the maximized expected reward of each action in a given state, integrating future consequences.</p>