Optimality theory of stigmergic collective information processing by chemotactic cells

arXiv:2407.15298 · q-bio.CB, nlin.AO · Submitted 2024-07-21 · Read on arXiv

Listen

Radio episode about this paper

Transcript

Introduction to the show: ident: Quantitative Biology. The day in quantitative biology: models, measurements and the data behind them.

Ines: We're looking at the paper: "Optimality theory of stigmergic collective information processing by chemotactic cells".

Title and authors: Ines: So, we're looking at this paper today, "Optimality theory of stigmergic collective information processing by chemotactic cells." It's about how groups of cells don't just follow a signal; they actively create their own signals to help them explore and find targets.

Marcus: Yeah, it tackles that idea of cooperation in cell populations where the functions they achieve are bigger than what any single cell could do on its own. It sets up this framework using reinforcement learning theory to figure out the best way for these agents to behave collectively.

Yuki: What I find interesting is how they frame collective exploration, where cells sense external cues from targets and then generate their own internal cues to coordinate that whole group's movement toward those targets. It moves beyond just individual sensing.

Ines: Exactly. The paper formulates this problem as a reinforcement learning problem, deriving the optimal exploration dynamics for these agents and comparing that structure to an established model called the Keller-Segel model, which is how we usually describe collective cellular dynamics in terms of gradients.

Marcus: It seems they find an optimal coupling between sensing that external gradient and generating their own internal gradient, but they point out this is different depending on whether the sensing mechanism itself is logarithmic or linear.

Yuki: That difference in the optimal way to generate that internal cue based on how they sense the external one sounds biologically very relevant, because cells have different ways of responding to chemical concentrations.

Ines: Right. The theory suggests that for a logarithmic sensing policy, the optimal way to generate an endogenous gradient involves production that is exponentially dependent on the exogenous cue concentration, whereas a linear sensing policy pairs with production that's linearly dependent on it.

Marcus: And what they show is that even though these two optimal couplings solve the same basic optimization problem, their representations of how those optimal values relate back to the internal state value are different.

Yuki: That suggests there might be other possible optimal couplings depending on some other mapping function, F, which is a very open door for biological interpretation.

Ines: They also use this RL structure to compare these collective dynamics against single-agent searching dynamics, and they find that distributed information processing by the population allows a fraction of agents to reach the target robustly.

Marcus: That comparison is telling; it suggests that even with limited computational power, if you have a population learning cooperatively through cues, you can achieve better overall results than one super smart agent might manage alone.

Yuki: It supports the idea of redundancy in biological systems; having many simpler agents working together can lead to a more robust response when dealing with environmental uncertainty.

Ines: They then take this and derive a continuous limit, which leads to partial differential equations that are structurally equivalent to the Keller-Segel models for collective chemotaxis. This bridges the gap between the learning dynamics and established physical models of cell movement.

Title and authors: Marcus: The continuous limit of their logarithmic coupling model corresponds to a diffusion-advection equation where the drift in space-time is dictated by how agents move along a random walk weighted by the exponent of their surrounding endogenous cue concentration.

Yuki: And for the linear coupling, they get a diffusion equation with linear degradation and linear production, which directly mirrors those established Keller-Segel models that describe chemotaxis.

Ines: So, they're showing that these two different sensing/generation strategies lead to structurally distinct continuous limits, even though both are biologically plausible descriptions of collective movement.

Marcus: The numerical experiments they did on a maze with two hundred seventy-four vertices showed that the agents collectively modify the endogenous cue concentration to reach the goal, expanding their exploration area by dispersing at branching vertices.

Yuki: They observed that after reaching the goal, the agents produce their own cue to reinforce that gradient toward the source of the exogenous cue, which eventually aligns from start to finish.

Ines: And when they compared optimal versus non-optimal couplings in a one-dimensional environment, they found that for the optimal log-exp and lin-lin couplings, the steady agent distributions are almost identical while their steady cue distributions are different.

Marcus: But for the non-optimal ones, like log-lin or lin-exp couplings, they saw dispersal and overaccumulation in those agent distributions.

Yuki: That really highlights how crucial it is to get that right coupling between sensing and modulation for achieving efficient exploration in a complex space.

Ines: They also compared this population learning with a single smart agent, and they found that the overall update speed of the internal cue by the single agent is slower than what happens in the larger population because of its limited exploration range.

Marcus: So, their conclusion here is that a single smart agent might not explore or learn as effectively as this smaller but distributed population does when it comes to searching for targets.

Yuki: That points toward an evolutionary advantage for simple organisms, suggesting that a population with limited computational capability can achieve more robust learning through distributed modulation of external cues than one smart agent with complete internal memory.

Ines: Overall, the paper "Optimality theory of stigmergic collective information processing by chemotactic cells" provides a foundation for understanding how cell populations can use dynamic sensing and modulation of cues to explore environments.

Marcus: It really shows that the optimal coupling between sensing and modulation isn't just one fixed thing; it depends on whether you choose logarithmic or linear gradient sensing.

Yuki: It gives us a way to think about how simple agents, like cells, can achieve complex collective behaviors by strategically manipulating their own chemical environment in response to external signals.

Ines: We're going to wrap up here for this paper. Keep an eye on how these continuous limits connect the learning dynamics to the established Keller-Segel models for chemotaxis.

The paper's summary: Ines: So, to recap this paper, they're using reinforcement learning theory to figure out how a group of cells can explore an environment by sensing chemical signals and then creating their own signals to coordinate the search toward targets.

Marcus: Right, it’s about that whole collective exploration idea where cells aren't just reacting blindly but are actively modulating their own chemistry to guide the whole population.

Yuki: From a population genetics standpoint, this is interesting because it suggests an evolutionary advantage for simple agents with limited computational power when they work together through cues.

Ines: Exactly. The paper identifies two main optimal ways this can happen depending on whether the cells sense the external cue using a logarithmic or a linear function.

Marcus: And that difference in sensing mechanism leads to different optimal ways they should generate their own internal cue, which is what makes this really deep theoretically, because it shows how biology dictates the best learning strategy.

Yuki: It connects back to those established models like the Keller-Segel model for chemotaxis, but they’re using a reinforcement learning structure to derive these dynamics, which is a new way to look at it.

Ines: That's right. They derive continuous limits from these two different coupling models, and they show that one limit is structurally equivalent to a diffusion-advection equation with logarithmic drift in space-time, while the other limit matches the Keller-Segel model exactly.

Marcus: So we’re looking at two distinct physical descriptions of collective movement emerging from the same RL framework.

Yuki: And their numerical experiments on a maze showed that for both optimal couplings, the agent distributions were almost identical, but how those internal cue concentrations were distributed was different based on which coupling they used.

Ines: It really highlights that getting the sensing and modulation right is crucial; a non-optimal coupling leads to dispersal and overaccumulation in the agent population.

Marcus: That makes sense because it shows the trade-off between maximizing the reward—reaching those targets—and controlling how fast you learn or move around.

Yuki: And comparing this to a single smart agent, they found that even with limited memory, a population learns more robustly than one highly intelligent cell.

Ines: So what this means for us is that distributed learning through chemical cues might be a key mechanism for how simple organisms achieve complex navigation in their environments.

Marcus: It points toward an evolutionary advantage where redundancy allows a population to overcome the limitations of any single agent’s internal memory or processing speed.

Yuki: It suggests that we need to model these collective dynamics not just as individual decisions, but as a whole system dynamically adjusting its own chemical landscape.

The paper's improvements: Tom: So, we're looking at what the authors suggest to improve this paper on stigmergic collective information processing by chemotactic cells today.

Ines: They basically say that while they found two optimal couplings, there are ways to extend this framework beyond just those two specific pairings.

Marcus: What do they mean by extending it? Are we talking about adding more biological complexity or changing the mathematical structure?

Ines: It’s about incorporating other known biological factors. For instance, they suggest you can derive learning dynamics that include a Hill function in the chemotactic response and production part.

Yuki: A Hill function? That would reflect the actual way receptors work in cells, which is a much more accurate description than just a simple linear or exponential relationship.

Marcus: And they also talk about modifying other parts of the RL framework, like changing the control penalty term or adding a regularization term.

Ines: They are exploring that using the principle of optimality to see what happens if you tweak things like how you map the cue-value and what kind of interactions between agents there are.

Yuki: That could open up a whole new area for population genetics because it lets us test how different types of genetic strategies, which might influence agent heterogeneity, affect collective exploration.

Marcus: So it’s not just about finding one perfect mathematical solution, but building a toolbox to explore many possible biological realities.

Ines: And they also touch on how you can modify the agent interactions themselves in this model.

Yuki: That would be interesting because in nature, agents don't always interact with every other agent; sometimes it’s very localized communication that matters most.

Marcus: It seems like the authors are pushing for a more biologically grounded model where the math reflects how real cells actually operate under various conditions.

Ines: They are trying to move from just an abstract RL mapping to something that directly models receptor kinetics and cellular behavior.

Yuki: That kind of detail is what population genetics needs; if we can map these dynamics onto different genetic architectures, we can make predictions about how species might evolve better ways to find food or avoid predators.

Conclusion: Tom: So we're wrapping up this session on the paper "Optimality theory of stigmergic collective information processing by chemotactic cells."

Ines: To summarize, they used reinforcement learning to find two optimal ways for cell populations to explore environments by balancing sensing and cue generation based on whether the sensing is logarithmic or linear.

Marcus: Right, it's about finding the best balance between reading the external chemical gradient and creating your own internal signal to guide everyone toward a target.

Yuki: It really shows that distributed information processing can be more robust than what any single agent can do on its own, which is important when you look at how species manage resources across a whole population.

Ines: Exactly. They found that the optimal coupling matters; choosing the right pairing between sensing and generation determines whether agents disperse or accumulate inefficiently.

Marcus: So the results show that those two optimal couplings, log-exp and lin-lin, give you the best outcome for reaching targets because they balance the reward component against a penalty for control cost.

Yuki: I think this means we can start thinking about how different genetic variants might influence which of these two coupling strategies is favored by a population in nature.

Ines: That’s right, and it opens up the door to understanding how simple agents can achieve complex navigation just by strategically manipulating their own chemical environment.

Marcus: It also shows that even with limited computational power, distributed learning through cues can outperform a single smart agent with full internal memory in search tasks.

Yuki: That’s a big idea for evolutionary biology; it suggests redundancy through distributed computation is an advantage for organisms with low replication costs.

Ines: Ultimately, the paper "Optimality theory of stigmergic collective information processing by chemotactic cells" gives us a foundation to study how cell populations use dynamic sensing and modulation of cues.

Marcus: It’s a solid piece of work that bridges reinforcement learning theory with established models like Keller-Segel for collective movement.

Yuki: It’s definitely something worth considering as we look at how different biological systems coordinate complex tasks through simple rules.

Department of Mathematical Informatics, the Graduate school of Information Science and Technology, the University of Tokyo · Institute of Industrial Science, the University of Tokyo · Universal Biology Institute, The University of Tokyo

q-bio.CB, nlin.AO

Submitted: 2024-07-21

Updated: 2025-01-02

Journal ref: PRX Life 3, 043005 (2025)

DOI: 10.1103/tvfy-lbbl

License: http://arxiv.org/licenses/nonexclusive-distrib/1.0/

The gist: Collective information processing by populations of cells is fundamental in various biological systems…

Key concepts

Collective Exploration Problem
This problem models how a group of agents (cells) cooperatively search for targets in an environment. Agents must learn policies to move and modify chemical cues simultaneously, where the goal is to maximize the accumulated reward, which is tied to finding high concentrations of an exogenous cue.
Optimal Coupling Relation
This refers to the best mathematical relationship between how agents sense existing external cues (like a gradient) and how they generate their own internal cues. The study proves this optimal coupling changes depending on whether the sensing mechanism is logarithmic or linear, dictating the most efficient way for the population to coordinate its behavior.
Keller-Segel Model Correspondence
The continuous mathematical limit of the derived cell dynamics matches well-known Keller-Segel models used to describe collective chemotaxis. This correspondence validates that the RL formulation successfully captures biologically relevant phenomena where cell movement is guided by chemical gradients generated by the cells themselves.

Terminology

Summary

Collective information processing by populations of cells is fundamental in various biological systems, where the cooperation of multiple cells results in complex functions beyond individual capabilities Collective information processing is fundamental in various biological systems, where the cooperation of multiple cells results in complex functions beyond individual capabilities A distinctive example is collective exploration where chemotactic cells not only sense the gradient of guiding exogeneous cues originating from targets but also generate and modulate endogenous cues to coordinate their collective behaviors In this study, we formulate the collective exploration problem as a reinforcement learning (RL) by a population Based on RL theory, we derive the optimal exploration dynamics of agents and identify their structural correspondence with the Keller-Segel model, the established phenomenological model of collective cellular dynamics Our theory identifies an optimal coupling relation between gradient sensing and gradient generation and demonstrates that the optimal way to generate a gradient qualitatively differs depending on whether the gradient sensing is logarithmic or linear The underlying RL structure is leveraged to compare the derived collective dynamics with single-agent searching dynamics, showing that distributed information processing by population enables a fraction of agents to reach the target robustly Our formulation provides a foundation for understanding the collective information processing mediated by dynamic sensing and modulation of cues

Formulation of Collective Exploration

The environmental exploration is represented as a connected undirected graph G = (S, E) with a finite set S of vertices and a finite set E of edges, which allows to incorporate general spatial topologies explored by the agent population The location of agent i ∈ 1, 2,... N at time t ∈ 0, 1,. p. 1.., is indicated by state s i t ∈ S Within a small unit time, agent i stochastically chooses either to stay at s' = s i t or to move to a neighboring vertex s' ∈ N (s i t) with probability π i t(s's i t) ∈ [0, 1] The transition probability π i t: S → ∆(S) is referred to as the agent’s policy Each agent is assumed to learn a better policy π i t to reach and accumulate at targets in the environment through exploration We consider a single exogenous cue for simplicity, whose spatial distribution is represented by r: S → R≥0 For a biological instance, r(s) = exp−l(s; s 0) represents fMLP concentration originating from damaged tissue at s 0, at distance l(s; s 0) from the tissue While exogenous cues may be affected by agents, we assume for simplicity that they are not altered by the agents and dependent only on the targets’ positions and magnitudes

Characterization of Optimal Exploration

To derive optimal degradation, production, and sensing of the endogenous cue for collective exploration, we associate the concentration of the exogenous cue r with the reward function in RL framework because reaching the targets is equivalent to reaching the place where the exogenous cue r is maximum Then, we formulate the problem as learning the optimal policy π∗ = argmaxπ J [π] that maximizes the following expected cumulative reward J [π] with discounting and control penalty: R π(s t, s t+1) = r(s t) − 1/β log π(s t+1s t) p(s t+1s t) The functional form of optimal policy means that the agent should bias its intrinsic transition p(s's) to the surrounding vertices with higher V1(s') This behavior can be interpreted as a biased random walk along the local gradient of V1(s') If agents encode Vt through the endogenous cue concentration Zt such that Vt(s) = FZt, they can update Vt(s) by modulating the endogenous cue via its degradation and production whilst moving by following the suboptimal policy with Vt in Eq. (4)

Optimal Coupling of Cue Sensing and Modulation

A biologically relevant mapping is the one associating the logarithm of the endogenous cue concentration with the estimate of the optimal state value: Vt(s) = FZt:= log(Zt(s))/β for s ∈ S Under this logarithmic mapping, the corresponding agents’ policy is reduced to s i t+1 ∼ π t(·s i t):= p(·s i t)Z γ t (·) P s'∈S p(s's i t)Z γ t (s') Meanwhile, the endogenous cue Zt should be modulated by its degradation and production for learning Zt which better approximates V∗ This learning dynamics is a distributed [52] and sequentially regularized version of greedy Z-learning [49], which have two advantages from both biological and learning theory perspectives

Continuous Limit of Optimal Cue Sensing and Modulation

To demonstrate the biological relevance of the policy (5) and cue dynamics (6), we derive its continuous limit, which is structurally equivalent to the wellestablished Keller-Segel models for collective chemotaxis [26–28] For simplicity, we consider a one-dimensional lattice S = S h:= 0, ±h, ±2h,... with a uniform spacing h and define the agents’ intrinsic transition probability p h: S h × S h → [0, 1] as a lazy random walk Under an appropriate time scale and in the limit of infinite agents, the continuous space-time limits h → 0 of the log-exp coupling models becomes the following pair of one-dimensional partial differential equations (7a) and (7b) Equation (7) is structurally homologous to the KellerSegel models [26–28]

Numerical Experiments

In this section, we conduct numerical experiments using the derived learning dynamics to evaluate their behaviors and optimality in biologically relevant problems We adopted a maze with S = 274 vertices shown in Fig. 2, which models the maze used in an experi The exogenous cue concentration r is assumed to be high only at a goal (target) vertex s goal ∈ S and uniformly low at the other vertices: r(s goal) = rtarget and r(s) = rdefault for s ≠ s goal Initially, the endogenous cue is uniformly distributed (Z0(s) = Z0 for s ∈ S), and all N agents start exploring from the vertex s i 0 = s start ∈ S Then, agents collectively modify the endogenous cue concentration to reach the goal s goal (Figure 2 a, b, and c) We observe that the agents expand their exploration area by dispersing at branching vertices and shaping a gradient toward unexplored directions via degradation of the endogenous cue (Fig. p. 5 2 a and b at t = 21 and t = 1001) Upon reaching s goal, they produce the endogenous cue, reinforcing the gradient toward s goal Eventually, the gradient aligns from s start to s goal and major fraction of agents accumulate near s goal (Fig. p. 5 2 a and b at t = 15001)

Coupling Optimality and its Consequence

To scrutinize how the optimality of coupling affects the agent’s behaviors, we compare the steady states in a simpler one-dimensional environment for the four couplings For the optimal couplings, i.e., log-exp and lin-lin couplings, the steady agent distributions µ T are almost identical while the steady cue distribution Z T are different, for each β (Fig. p. 6 3 c) However, for the non-optimal couplings, i. p. 6e., log-lin and lin-exp couplings, dispersal and overaccumulation in µ T were observed These results highlight the importance of optimal coupling between cue sensing and modulation for achieving efficient exploration and navigation in a complex environment

Comparison with a Single-Agent System

We have analyzed distributed learning in which agents with limited memory share and learn environmental information cooperatively through the endogenous cue to search for targets We address this problem by leveraging the fact that both of our distributed model and the usual single-agent learning model are based on the same RL framework In this one-trial comparison, the overall update speed of Zt by the single agent is slower than the agent population due to its limited exploration range (Fig. 4 a t = 1001) These observations suggest that the single smart agent may not explore and learn as effectively as the less smart agent population

Summary and Discussion

To formulate a cell population that distributedly explores and learns target information via chemical cues in the environment, we leveraged reinforcement learning theory and derived two qualitatively different optimal dynamics of the population with linear and logarithmic gradient sensing These dynamics have continuous limits that are structurally equivalent to the phenomenological Keller-Segel model, indicating that optimal learning could be achieved by biologically relevant cell populations We also verified that the optimal coupling of the sensing and modulation of cues is crucial for achieving efficient exploration by the population Additionally, we have demonstrated that an agent population with limited computational capability can achieve more robust learning via distributed modulation of external cues than a smart single agent with complete internal memory

Extensions of the Formulation

Although we derived only two optimal couplings, our formulation can be extended beyond them by incorporating various biological factors, providing a foundation for the quantitative understanding of collective information processing mediated by extracellular signals For instance, we can derive a learning dynamics with a Hill function in chemotactic response and production, reflecting the underlying receptor kinetics [27–35] Furthermore, we may explore various collective dynamics from the principle of optimality by modifying (i) the cue-value mapping and control penalty, (ii) regularization term, and (iii) agent’s interactions and heterogeneity

Evolutionary Advantages of Distributed Computation

Our result demonstrates that a population of simple agents can excel at learning target information robustly compared to a single smart agent This aligns with the redundancy principle [77], which states that having many particles with the same objective can increase the probability of quickly achieving the objective as a population According to this biophysical principle, distributed information processing may have an evolutionary benefit for agents with limited capabilities but low replication costs For example, neutrophils may achieve a robust population response by dispersing numerous genetically identical cells

Pruning redundant reasoning tokens

The paper demonstrates that the optimal endogenous cue modulation qualitatively differs depending on whether the gradient sensing is logarithmic or linear The logarithmic sensing policy optimally couples with production that is exponentially dependent on exogenous cue concentration, whereas the linear sensing policy optimally couples with production that is linearly dependent on the concentration These two are almost identical in terms of the optimization criterion, and there may exist other optimal couplings, corresponding to different choices of the mapping F[Z]

How the training signal is built

The log-exp coupling optimally couples with production that is exponentially dependent on exogenous cue concentration, whereas the linear sensing policy optimally couples with production that is linearly dependent on the concentration The log-exp coupling and lin-lin couplings are structurally equivalent in terms of solving the optimization problem The optimal couplings solve the same optimization problem, while their representations of V via Z are different

Where the gains come from

The optimal log-exp and lin-lin couplings exhibit almost identical and higher values V π compared to the non-optimal log-lin and lin-exp couplings at any β The optimal couplings achieve a good balance between the reward component V π reward the penalty component V π penalty/β This consistency at both s 0 and s 19 reinforces the robustness of these trends across vertices The deficiency of the non-optimal couplings is more pronounced at s 0 than at s 19, especially for the non-optimal lin-exp coupling: for large β, the lin-exp coupling incurs a much higher penalty V π penalty(s 0) (Fig. S2 a), resulting in a significantly lower V π(s 0) (Fig. p. 5 p. 5 p. 5 S2 b) This discrepancy arises because V π reward(s) places more emphasis on rewards near the starting vertex s, which diminish as s are farther from the goal s 19 The results demonstrate that non-optimal couplings are indeed inferior in terms of optimization for accumulating at targets, highlighting their qualitative difference in behavior

The optimal log-exp coupling and lin-lin couplings exhibit almost identical and higher values V π compared to the non-optimal log-lin and lin-exp couplings at any β The optimal couplings achieve a good balance between the reward component V π reward and the penalty component V π penalty/β This is a natural consequence of the fact that the two optimal couplings solve the same optimization problem, while their representations of V via Z are different The optimal couplings solve the same optimization problem, while their representations of V via Z are different The optimal couplings solve the same optimization problem, while their representations of V via Z are different

The optimal log-exp and lin-lin couplings exhibit almost identical and higher values V π compared to the non-optimal log-lin and lin-exp couplings at any β The optimal couplings achieve a good balance between the reward component V π reward and the penalty component V π penalty/β This consistency at both s 0 and s 19 reinforces the robustness of these trends across vertices The deficiency of the non-optimal couplings is more pronounced at s 0 than at s 19, especially for the non-optimal lin-exp coupling: for large β, the lin-exp coupling incurs a much higher penalty V π penalty(s 0) (Fig. S2 a), resulting in a significantly lower V π(s 0) (Fig. S2 b) This discrepancy arises because V π reward(s) places more emphasis on rewards near the starting vertex s, which diminish as s are farther from the goal s 19 The results demonstrate that non-optimal couplings are indeed inferior in terms of optimization for accumulating at targets, highlighting their qualitative difference in behavior

The optimal log-exp and lin-lin couplings exhibit almost identical and higher values V π compared to the non-optimal log-lin and lin-exp couplings at any β The optimal couplings achieve a good balance between the reward component V π reward and the penalty component V π penalty/β This consistency at both s 0 and s 19 reinforces the robustness of these trends across vertices The deficiency of the non-optimal couplings is more pronounced at s 0 than at s 19, especially for the non-optimal lin-exp coupling: for large β, the lin-exp coupling incurs a much higher penalty V π penalty(s 0) (Fig. S2 a), resulting in a significantly lower V π(s 0) (Fig.

Green — the paper says it, and the page is printed. Red — no passage carries it, so treat that sentence as unverified. The same check runs on your own text: check a summary against its source, $5.

Improvements for AI systems

  1. Optimal learning dynamics for chemotactic cells can be derived using reinforcement learning theory, allowing for distributed information processing by a population of agents to reach targets robustly compared to a single smart agent with complete internal memory.

  2. The formulation provides a foundation for understanding collective information processing mediated by dynamic sensing and modulation of cues, identifying an optimal coupling relation between gradient sensing and gradient generation that differs based on whether the sensing is logarithmic or linear.

  3. The continuous limit of the log-exp coupling model corresponds to a diffusion-advection equation with the logarithmic gradient as drift in space-time, which models how agents move along a random walk weighted by the exponent of their surrounding endogenous cue concentration.

  4. The continuous limit of the lin-lin coupling model corresponds to a diffusion equation with linear degradation and linear production, providing a structural correspondence to established Keller-Segel models for collective chemotaxis.

  5. The comparison between optimal and non-optimal couplings demonstrates that while the non-optimal lin-exp coupling can achieve comparable timeaveraged reward Ri, it incurs a trade-off in temporal control costs, highlighting the difference between optimizing for Ri versus V pi.

  6. The analysis shows that an agent population with limited computational capability can achieve more robust learning via distributed modulation of external cues than a smart single agent with complete internal memory, suggesting an evolutionary advantage for simple and fertile organisms.

Sources

Related papers