Environmental requirements for the use of social information by artificial life agents using evolved plastic artificial neural networks

Hugh Charterton; James M. Borg; Aniko Ekart

cs.AISep 28, 2026

Environmental requirements for the use of social information by artificial life agents using evolved plastic artificial neural networks

Authors: Hugh Charterton, James M. Borg, Aniko Ekart

Organizations: Aston Centre for AI Research and Application, Aston University, Birmingham, United Kingdom

Abstract

Evolved Plastic Artificial Neural Networks (EPANNs) consist of two principal processes, the first, evolution, and the second, development and in-life learning. In the context of the origins of social = learning, very few studies have been carried out using ALIFE models based on EPANN requirements. Studies in this field have usually involved an imitative teacher/pupil relationship. This, however, ignores the possibility that the observed behaviour is a consequence of social information cues rather than direct imitation or teaching. Starting with the first of the EPANN processes (evolution), a series of experiments was undertaken using artificial neural network (ANN) based agents in a variety of foraging environments to examine under what minimal environmental conditions the use of social information might have evolved, as measured by the number of generations taken to meet a specified fitness criterion. NEAT (Neuroevolution of Augmenting Topologies) was the ANN used as its evolutionary algorithm would evolve a network's topology as well its weights. Unintentionally, in the experiment there was a simple network topology based on the location of the nearest food item which enabled agents to swiftly meet the fitness criterion. With this topology, additional information, social or otherwise, was not required and could have proved to be a hindrance. However, this does indicate that for the use of social information to have evolved, it would require a greater degree of complexity in the environment to do so.

Figures & tables

Configuration Item	Description and setting
pop_size	Number of agents (dynamically set as per the run requirements).

num_inputs	Number of network inputs (see table 5 for values).

feed_forward	Network configuration. Set to false to enable recurrent connectivity.

Table 1: NEAT configuration values (where different from default).

Environment Variables	Value a	Value b	Value c	Value d	Value e
Agent Population	30	40	50	60	70

Total Food/Poison Items	40	45	50	55	60

Food/Poison Ratio	1:1	1.3:1	1.8:1	2:1

Table 2: Variable values for agent population, total number of food/poison items and the varying food/poison ratios. During a run, additional food/poison items would re-spawn if a random number between zero and one was less than 0.2 times the food/poison ratio.

Input Code	Description	# NEAT Inputs
FP_XY	XY co-ordinates of the nearest visible Food/Poison item relative to agent.	2

FP_IND	Indicator of whether an item is either food or poison.	1

AG_EN	Normalized value (n) of agent’s own energy (e) as calculated in equation 1 $n=max(0.0,min(1.0,e/f))$ (1) where $f$ refers to fitness threshold in units of energy	1

Table 3: Inputs, descriptions and the number of NEAT inputs required.

Interaction Id	direct/indirect	Description
1	direct	Information about item and the agent its self.

2	indirect	Information about another agent.

3	direct & indirect	Combined direct and indirect information

Table 4: Direct/Indirect Interaction types.

NEAT input combination ID	Inter-action ID	Input codes	Total Inputs
1	1	FP_XY	2
2	1	FP_XY, FP_IND	3
3	1	FP_XY, AG_EN	3
4	1	FP_XY, FP_IND, AG_EN	4
5	2	NA_XY, NA_B	3
6	2	NA_XY, NA_B, AG_EN	4

Table 5: NEAT input combinations, grouped by their type of interaction with the environment and their total number of inputs to a NEAT neural network. Please note that combination 6 has been classified as indirect since the agent’s energy is viewed as complementary to the information about another agent.

Output code	Description	# Outputs
X	Movement along X axis	1
Y	Movement along Y axis	1

Table 6: NEAT neural network output values. As NEAT would return values between -1 and 1, a speed value was used to map the returned value to give the number of pixels to move. The speed value was set to 5.

Figure 1: Number of fitness thresholds met for each ratio of food/poison items for each initial total of food/poison items in the environment across all populations and all input code combinations.

Figure 2: Number of fitness thresholds met across the range of population values for each initial total of food/poison items in the environment across all food/poison ratios and all input code combinations.

Figure 3: Fitness thresholds meeting the termination criteria for each NEAT input combination. For NEAT input combinations, see table 5 .

Figure 4: Percentage of thresholds met in their first generation out of the total number of thresholds met per input code combination. where 9-11% of each of the direct NEAT input combinations meets the threshold in their first generation, a strong indicator that this environment might have been too simple. Those with combined direct and indirect NEAT input combinations meet the threshold approximately 1% of the time in their first generation, while the indirect combinations never reached the threshold in their first generation. For details on each input code combination, see table 5 .

Panel of Figure 5 — Figure 5: Examples of first generation NEAT topologies which reached the fitness threshold in which NEAT inputs connect directly to the outputs with no hidden layers. Both examples had initial agent populations of 30, initial number of food/poison items of 55 with a ratio of $2:1$ .

Figure 6: Cumulative number of fitness thresholds met per input code combination as defined in table 5 .

Explore similar work

SAGE: A Quantitative Evaluation of Socialized Evolution in Agent Ecosystems
Jun 2, 2026Linyue Pan, Yaoming Zhu, Lin Qiu +2Self-Evolving Agents Self-Improving Agents
Tsinghua University, China · Meituan, China
Neuroevolution Arena: Nested Ecological Evaluation of Update-and-Inheritance Regimes across Neural Architectures
Aug 10, 2026Yuxu Ge, Yifei ChengNeuroevolution Neural Architecture Search
University of York, United Kingdom
Emergence of agriculture in an artificial society of reinforcement learning agents
May 21, 2026Gautier Hamon, Martí Sánchez-Fibla, Clément Moulin-Frier +1Emergence Ecosystems
Flowers AI and CogSci lab, Inria, Université de Bordeaux, France · Department of Information and Communications Technologies, Universitat Pompeu Fabra, 08018, Barcelona, Spain · Artificial Intelligence Research Institute, IIIA (CSIC), Campus de la UAB, Bellaterra, Barcelona, 08193, Spain +4

Jun 2, 2026, cs.AIJ/K move · Enter open · S save

SAGE: A Quantitative Evaluation of Socialized Evolution in Agent Ecosystems

Linyue Pan, Yaoming Zhu, Lin Qiu, Xuezhi Cao+1

Tsinghua University, China · Meituan, China

Self-Evolving Agents Self-Improving Agents Self-Evolution Alphaevolve Depth History

Open paper AskPDF arXiv

Self-improving language agents are typically evaluated in isolation: an agent attempts a task, receives feedback, and iteratively refines its own behavior. Yet agents increasingly operate alongside peers whose strategies and outcomes are publicly visible. This raises an under-studied question: when does shared experience produce improvements that self-improvement alone cannot achieve? We introduce SAGE (Social Agent Group Evolution),an evaluation framework that compares two compute-matched conditions: SocialEvo, where agents from five distinct model families co-evolve with access to all peers' histories; and SelfEvo, where each agent receives the same number of task attempts but sees only its own past, which is conventional in self-improving agent studies. We instantiate SAGE in three arenas: open-ended ML research, long-horizon economic planning, and strategic multiplayer play, evaluated across multiple evolutionary rounds. We find that group history is not a universal amplifier: the strongest agent does not exceed its self-evolution ceiling. However, agents that plateau under self-improvement can achieve significant breakthroughs when peer experience is available. In competitive settings, counterfactual controls reveal that agents improve generally rather than developing opponent-specific strategies. Across different forms of shared history, filtered peer traces and reflective summaries often outperform raw logs, indicating that social gains depend on abstraction rather than exposure volume. These findings reveal that peer-history gains are agent-specific, arena-dependent, and contingent on the capacity to abstract transferable knowledge from public traces.

Abstract

Explore similar work

SAGE: A Quantitative Evaluation of Socialized Evolution in Agent Ecosystems

Neuroevolution Arena: Nested Ecological Evaluation of Update-and-Inheritance Regimes across Neural Architectures

Emergence of agriculture in an artificial society of reinforcement learning agents