Drive competition underlies effective allostatic orchestration

Abstract

Living systems ensure their fitness by self-regulating. The optimal matching of their behavior to the opportunities and demands of the ever-changing natural environment is crucial for satisfying physiological and cognitive needs. Although homeostasis has explained how organisms maintain their internal states within a desirable range, the problem of orchestrating different homeostatic systems has not been fully explained yet. In the present paper, we argue that attractor dynamics emerge from the competitive relation of internal drives, resulting in the effective regulation of adaptive behaviors. To test this hypothesis, we develop a biologically-grounded attractor model of allostatic orchestration that is embedded into a synthetic agent. Results show that the resultant neural mass model allows the agent to reproduce the navigational patterns of a rodent in an open field. Moreover, when exploring the robustness of our model in a dynamically changing environment, the synthetic agent pursues the stability of the self, being its internal states dependent on environmental opportunities to satisfy its needs. Finally, we elaborate on the benefits of resetting the model's dynamics after drive-completion behaviors. Altogether, our studies suggest that the neural mass allostatic model adequately reproduces self-regulatory dynamics while overcoming the limitations of previous models.

Keywords: allostatic control; attractor model; control theory; homeostasis; need-based behavior; self-regulation.

FIGURE 1

Performance of the neural mass allostatic model in an open field. Two gradients represent the areas where the two internal needs, arousal (red) and security (green), can be fulfilled (top-left). The agent partially observes those gradients through local sensation (top-middle). Local sensation allows the agent to adjust its actions to ascend/descend the gradients and detect when the observer’s current position is in the vicinity of the resource area (i.e., the peak of the gradient). If the agent is not close enough to the resource area, the internal state related to that resource keeps declining, as the security actual state (aS) is doing in this figure (top-right). In contrast, if the agent occupies the peak of the gradient, the internal state approximates the desired state (dS), as shown with arousal. The aSs and dSs are compared, creating a homeostatic error (top-right) that will input their respective excitatory pools in the neural mass model (middle-right). The level of competition is defined by the ratio of mutual inhibition (orange interneurons) and shared feedback inhibition (blue interneurons). Finally, the firing rate of each excitatory population provides the agent with the corresponding drives given its internal states.

FIGURE 2

Gradients representing environmental opportunities to satisfy internal needs in each condition. (A) Arousal and security gradients were designed to replicate rodent behavior in an open field test. To represent the maximal level of exposure when exploring the center of the arena, we set the peak of the gradient in that location. Similarly, we used the peak of the security gradient to represent the home base location in one of the corners. (B) Hydration and temperature gradients were designed to test the performance of our model in a dynamic setting. Here, the hydration gradient was static, with its peak in one of the corners. In contrast, the temperature gradient changed over time. To do so, we built the gradient as a two-dimensional sigmoidal function where its x-intercept increased as the simulation evolved. Thus, the peak area where internal temperature increases shrunk, and the intermediate area between gradients increased over time.

FIGURE 3

Replication of rodent behavior in an open field test. (A) Agent’s trajectory tracked along a complete experiment. The red dot represents starting location, randomized across experiments. (B) Agent’s internal state dynamics during a complete experiment (same as a). (C) Firing rates of the excitatory populations configuring the neural mass allostatic model during a complete experiment (same as a). (D) Occupancy map across 50 experiments with normalized values. (E) Distribution of security states across 50 experiments. (F) Distribution of arousal states across 50 experiments. (G) Mean attractor dominance across 50 experiments.

FIGURE 4

Agent’s performance in a dynamic environment. (A) Agent’s trajectory tracked along a complete experiment. The red dot represents starting location, randomized across experiments. (B) Agent’s trajectory tracked along a complete experiment (same as a) divided into five periods. (C) Map of occupancy divided into five periods. Color crosses illustrate the mean position during hydration (blue) and thermoregulation (orange) attractor dominance. The red dashed line illustrates the variable slope location of the dynamic temperature gradient. (D) Mean attractor dominance (E) Mean agent internal state dynamics and environmental temperature. The shaded area indicates the internal state variance. (F) Mean internal state dynamics correlated with the environmental temperature. (G) Efficiency dynamics correlated with the environmental temperature. (H) Fairness dynamics correlated with the environmental temperature. (I) Stability dynamics correlated with the environmental temperature. Aggregated results from 50 experiments.

FIGURE 5

Decision-reset supports the agent’s internal stability. (A) Comparison of temperature and hydration internal states and efficiency, fairness, and stability scores between no-inhibition (study 2) and decision-reset (study 3) conditions. (B) Correlation between internal states and game theory measures with environmental temperature for both no inhibition and decision-reset conditions.