Exploration Behavior of Untrained Policies

The Big Picture

Imagine being dropped into an unfamiliar city with a mostly blank map. How you explore (wandering randomly, walking in straight lines, gridding the streets) determines how quickly you fill it in. AI agents trained through trial and error, known as reinforcement learning or RL, face the same challenge: they must explore their environment to learn, but before learning anything, they have no idea where to go.

Most RL research focuses on rewards: bonus points for visiting new places, penalties for getting stuck. But Jacob Adamczyk at UMass Boston and IAIFI asks a more basic question: what if the underlying structure of an AI’s decision-making network, before any training begins, already determines how an agent explores? What if the shape of that decision network encodes hidden exploration biases nobody has been paying attention to?

The answer is yes. And it changes how we should think about exploration.

Key Insight: Untrained neural network policies aren’t random noise generators. Their architecture imposes structure on exploration trajectories, producing either smooth “ballistic” motion or diffusive random walks depending on how the policy is initialized and used.

How It Works

The core observation: a neural network with randomly initialized weights is not a random function. Neural networks are smooth. They map nearby inputs to nearby outputs. In RL, this smoothness has a direct physical consequence.

Two distinct trajectory regimes emerge from different ways of using an untrained policy:

• Ballistic trajectories: near-linear paths through state space, produced when an agent uses the same untrained policy throughout an episode
• Diffusive trajectories: broadly spreading, heavy-tailed distributions, produced by resampling a fresh random policy at every timestep

Ballistic trajectories arise because a smooth network maps similar states to similar actions. With continuous environment dynamics, the agent drifts like a ball rolling down a hill. Adamczyk formalizes this with a Lipschitz continuity argument: if the policy is $L_\pi$-Lipschitz and the environment dynamics are similarly well-behaved, the trajectory deviates from a straight line by at most an error that grows quadratically in time. Standard multilayer perceptron networks (MLPs with two hidden layers, 256 nodes, ReLU activations) produce exactly this behavior in simulation.

Diffusive trajectories require a different strategy: resampling a fresh random policy at every timestep. This creates a stochastic trajectory ensemble even with fully deterministic dynamics. The math becomes tractable through the infinite-width limit. As a network’s hidden layers grow arbitrarily wide, the distribution over functions it represents converges to a Gaussian Process (GP), a mathematical object that describes families of smooth random functions.

This is where physics enters. Once the action sequence is treated as draws from a GP, the agent’s trajectory becomes a stochastic process you can analyze with tools from statistical mechanics. The Fokker-Planck equation, which governs how probability distributions evolve over time, describes the dynamics. The kernel structure of the GP (determined by network architecture and activation functions) sets the diffusion coefficient and hence the character of exploration.

Resampled policies produce heavy-tailed steady-state distributions over visited states, so the agent spreads broadly rather than clustering near its start. Ballistic and diffusive regimes aren’t mutually exclusive; combining them offers a design knob for controlling the geometry of exploration before any training begins.

Why It Matters

Exploration becomes a problem of architecture design, not reward shaping. Most existing exploration methods (curiosity bonuses, count-based rewards, Thompson sampling) require extra networks, modified reward signals, or computational overhead. Adamczyk shows you can get meaningful control over exploration for free, simply by choosing how you initialize and resample your policy before training starts.

The connection to physics is more than decorative. Fokker-Planck analysis and GP theory are well-established in statistical mechanics and quantum field theory, and importing them into RL opens up new ways to analyze exploration rigorously. Heavy-tailed distributions from diffusive trajectories resemble Lévy flights, a pattern found in animal foraging and long studied as an optimal exploration strategy in uncertain environments. That the same pattern emerges naturally from neural network priors is a nontrivial result.

The toy model here deliberately sidesteps complexities of real RL: value-based algorithms, stochastic environments, high-dimensional state spaces. But the theoretical scaffolding (Lipschitz bounds, GP limits, Fokker-Planck analysis) should translate to richer settings. One could even imagine a new role for neural architecture search: rather than optimizing purely for expressivity or generalization, search for architectures whose initialization distributions produce favorable exploration geometries for a given task.

Bottom Line: Untrained neural network policies already encode rich exploration behavior through their architecture. By understanding this structure theoretically, researchers can treat policy initialization itself as a design tool, independent of rewards or auxiliary training machinery.