How does incorporating realistic individual cognition into models change predictions about collective animal motion compared with simple particle-force models?

Why treating animals as decision-makers changes the model

Simple particle-force models describe each animal as a self-propelled particle pushed and pulled by social forces such as short-range repulsion and longer-range attraction or alignment [2]. That picture is useful, but it treats an organism as a particle rather than a probabilistic decision-maker, and it leaves out the discrete, random, and internal-state-dependent parts of behavior that collective-motion models typically fail to include [1][2].

A cognitive approach replaces or supplements those built-in rules with an internal process. In an active-inference model, behavior follows from a single imperative, minimizing surprise, and agents act to suppress prediction errors that conflict with their expectations [2]. Framed this way, the classical social forces can be recovered and generalized as consequences of that error-suppression, rather than being written into the model as fixed rules [2]. A probabilistic framework takes a different route to the same gap, letting each animal switch stochastically between behavioral states, with each state producing a possibly different law of motion through space [1].

What the cognitive models actually predict about the group

The most concrete change is that large-scale collective patterns can appear without being built in. Cohesion, milling, and directed motion emerge naturally in the active-inference model even though no behavioral rules or goals are explicitly programmed into the individual agents [2]. That matters because a particle-force model typically produces these patterns by design, through its attraction and alignment terms; the cognitive model instead generates them from the agents' belief-driven attempts to reduce surprise [2].

Cognition also changes how individual differences map onto group properties. Exploring the belief-based model's parameter space revealed nontrivial relationships between individual beliefs and group-level features such as polarization and the tendency to visit different collective states, and individual beliefs about uncertainty were found to shape how accurately the group makes collective decisions [2]. The model also allows agents to update their generative model over time, which produced groups that were collectively more sensitive to external fluctuations and encoded information more robustly [2]. These are simulation results, so they describe what the model predicts rather than what has been measured in a real school or flock.

Behavioral states and randomness in real fish

The probabilistic framework connects internal state to motion in a way that can be fitted to tracking data. Each animal switches between behavioral states at rates that can depend on the environment and on interactions with other animals, and the switching dynamics are represented as a generalized linear model whose rate functions can be inferred from data by gradient descent [1]. The authors applied the framework to two synthetic motion models and to real zebrafish tracking data, showing that forward simulation of multiple interacting animals is possible and that model selection can identify which factors modulate state switching [1].

A separate fish study points to a related balance without invoking internal states. From roughly 600 hours of video, researchers extracted a force map of fish interactions from trajectories using a convolutional neural network, interpreting that force as the fish's perception of surrounding individuals, the environment, and social information [3]. The fish were mostly in a seemingly disordered swarm state, yet their local interactions were clearly specific, and combining those interactions with the fish's inherent stochasticity reproduced the observed collective motions in simulation [3]. The authors concluded that a delicate balance between specific local force and intrinsic stochasticity is essential for ordered movement [3]. This supports the idea that randomness and local response matter, but it is a force-and-noise account rather than a test of cognitive states.

What these models have not yet settled

None of these studies directly compares the predictive accuracy of a cognitive model against a simple particle-force model on the same data. The probabilistic framework paper demonstrates how individual behavioral states influence collective dynamics but does not provide that direct comparison [1], and the active-inference paper discusses classical particle-force models and shows that social forces can be recovered within its framework, but it does not directly compare predictive outcomes between the two approaches [2]. The fish study likewise does not directly compare its simulations with predictions from simple particle-force models [3].

So the honest answer is that cognition changes the kind of explanation available and expands what the models can generate, including state-dependent switching, belief-driven pattern formation, and belief-dependent collective decision accuracy [1][2]. Whether those additions make forecasts about real animal groups more accurate than particle-force models is a question these studies raise but do not yet answer [1][2][3].

Sources used in this answer

Related research questions