Source-linked AI summary
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
TL;DR
The paper examines whether social information can evolve in artificial-life agents, addressing work that often frames social learning as imitation rather than responses to social cues. It tests NEAT-based agents in varied foraging environments and finds that simpler direct food-location solutions can preclude a clear test of social information; the study leaves environmental complexity unresolved.
Problem
Social-learning studies often use imitation or teaching, while the possibility that behavior responds to social information cues is also raised.
Source: Introduction · p. 1Method
The study used NEAT-evolved foraging agents with different combinations of direct environmental and indirect social inputs, without adding within-lifetime plasticity.
Source: Technical Background · p. 2, Results · p. 4Results
10% of the direct input combinations had a solution before NEAT's evolutionary algorithm was invoked, and simpler survival mechanisms made social information unlikely to emerge in this experiment.
Source: Conclusion · p. 6Takeaways & Limitations
The findings indicate that social information is unlikely to evolve when simpler survival mechanisms are available; the environmental complexity needed remains unexplored.
Source: Conclusion · p. 6The study tested only the evolutionary component of an EPANN, without within-lifetime plasticity.
Source: Technical Background · p. 2, Conclusion · p. 5
Abstract
from arXiv · showhide
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.
Introduction
The paper asks what environmental conditions might promote plasticity, learning, and intelligence, and focuses on social learning beyond direct imitation. Prior ALIFE models often use teacher–pupil transfer, while social cues may also guide behavior without teaching.
Introduction
- The paper raises what environmental conditions might promote the evolution of neural plasticity, learning, and intelligence.It situates this question within EPANN research and related work on plasticity, evolutionary robotics, evolved learning, and neuromodulation.Source: Introduction · p. 1
- The paper frames social learning as learning about the environment through communication, observation, or imitation of other EPANNs.This was proposed as a possible role for incremental learning.Source: Introduction · p. 1
- Earlier ALIFE social-learning models commonly transfer knowledge through imitation in teacher–pupil or parent–child relationships.The cited discussion notes that knowledge in these relationships must ultimately come from experiential learning.Source: Introduction · p. 1
- Social cues can guide a response without direct imitation; the paper cites monkeys responding fearfully after hearing another monkey's fearful response to a snake.The passage attributes the response to recognizing a fear cue, rather than copying the experienced monkey's behavior.Source: Introduction · p. 1
Technical Background
NEAT can evolve and grow diverse network structures through mutation and recombination. In this study, however, it was used without the within-lifetime plasticity needed to support emergent learning.
Technical Background
- NEAT can mutate and recombine network structures, including nodes, connections, layers, and weights, allowing diverse structures to evolve and grow.It is cited as an example of an EPANN evolutionary algorithm.Source: Technical Background · p. 2
- NEAT was used without adding the within-lifetime plasticity required to support emergent learning.The paper therefore uses NEAT as an evolutionary algorithm, not as a complete EPANN with in-life learning.Source: Technical Background · p. 2
- The experiment used NEAT-python, whose implementation is documented as differing from the original NEAT.Reproduction occurs at the end of a generational cycle, using some fittest members and a random selection of other individuals.Source: Technical Background · p. 2
- The NEAT configuration generally followed default settings, with exceptions listed in Table 1.The supplied table caption identifies the table as NEAT configuration values that differ from default.Source: Technical Background · p. 2, Table 1 · p. 2
Experimental Setup
The experiment varied agents and food/poison conditions in a foraging environment, using NEAT inputs that could represent environmental items, other agents, and the agent itself. Agents acted through two movement outputs, and runs tested combinations of these inputs and environmental settings.
Experimental Setup
- NEAT-based agents and varying numbers of food and poison items were placed in an 800 × 600-pixel rectangular foraging environment.The food/poison totals and ratios varied; additional items respawned under the probability condition stated in Table 2.Source: Experimental Setup · p. 2, Table 2 · p. 3
- Agents began with 100 energy; consuming food or poison changed energy by +50 or −50 and fitness by +10 or −10, respectively.Energy tracked mortality separately from fitness, which tracked progress toward the threshold; an agent died when energy reached zero.Source: Experimental Setup · p. 2
- A beacon reflected an agent's energy and could offer others indirect social information about its foraging health.Only information within an agent's visibility range—one quarter of the environment's diagonal—was available as input.Source: Experimental Setup · p. 3
- Each NEAT network had two outputs to update agent position, and 20 runs of 120 generations with 1,000 timesteps per generation were executed for each configuration.At each generation's end, the experiment logged initial variables, fitness-threshold outcome, and the fittest agent's network structure.Source: Experimental Setup · p. 3
- The NEAT inputs represented nearby food or poison coordinates, item identity, and the agent's own normalized energy.The input descriptions and required input counts are identified in Table 3.Source: Experimental Setup · p. 3, Table 3 · p. 3
- Agents received direct environmental inputs most often met the fitness threshold, combined direct and indirect inputs ranked next, and indirect-only inputs hardly ever met it.Runs covered combinations of the environment variables; the fitness-threshold outcome was logged at each generation.Source: Experimental Setup · p. 3
Results
Threshold success increased with the food-to-poison ratio, while larger populations showed an overall decline in effectiveness. Direct-input combinations performed best, and indirect-only combinations rarely or never met the threshold, consistent with a simple food-location topology that could succeed without social information.
Results
- Thresholds were met more often as the food-to-poison ratio increased, with similar counts across initial food-and-poison totals.Figure 1 summarizes counts across populations and input combinations.Source: Results · p. 3, Figure 1 · p. 4
- Population variation had no discernible impact apart from a steady overall effectiveness decrease as population grew, despite higher item volumes.Figure 2 presents threshold counts across population values and initial item totals.Source: Results · pp. 3–4, Figure 2 · p. 4
- Direct-input combinations met the most fitness thresholds, combined direct-and-indirect combinations ranked second, and indirect-only combinations hardly ever succeeded.The extended indirect-only runs met no thresholds.Source: Results · p. 4, Figure 3 · p. 5
- Between 9% and 11% of successful direct-input runs met the threshold in generation one, using direct connections from nearest-item coordinates to the outputs.The example networks had no hidden layers.Source: Results · p. 5, Figure 4 · p. 5, Figure 5 · p. 5
- Within each interaction type, the leading input combination used the fewest inputs; cumulative threshold counts rose faster initially for direct inputs than for combined inputs.Indirect combinations hardly registered in the cumulative results.Source: Results · p. 5, Figure 6 · p. 5
Conclusion
The experiment did not establish how social information use evolves, but suggests it is unlikely to emerge when simpler survival mechanisms are available. The authors identify environmental complexity and the experiment’s evolutionary-only, short-timescale scope as important boundaries.
Conclusion
- The study tested only EPANN evolution, across short generational runs in environments where items changed only through consumption and respawning.Development and in-life learning were not included.Source: Conclusion · pp. 5–6
- 10% of direct input combinations had solutions before NEAT evolution, and simpler survival mechanisms may prevent social information use from emerging.The authors conclude the environment was too simple to show whether social information improves foraging performance.Source: Conclusion · p. 6
- Food/poison impact was not varied, leaving nearest-item coordinates as a possible easy solution to survival.The authors suggest varying item impact and other environmental features to increase complexity.Source: Conclusion · p. 6
- Proposed environmental changes include consumption choices, agent reactions, varied item effects, clustered items, and items that persist beyond one timestep.Seasonal availability over a longer generational timespan is another proposed change.Source: Conclusion · p. 6
- Further work could log agent movement and proximity, then test whether social-information inputs evolve in a replicated Borg and Channon environment.The authors also propose longer runs with seasonal item availability.Source: Conclusion · p. 6
- The passive energy beacon may be public information; controlling its value could make it social information, with decoding potentially supplied by within-lifetime learning.The authors identify neuromodulated learning as the other half of an EPANN.Source: Conclusion · p. 6