[Paper Review] Increasing Behavioral Complexity for Evolved Virtual Creatures with the ESP Method

Dan Lessin, Don Fussell+2 more|arXiv (Cornell University)|Oct 27, 2015

Artificial Intelligence in Games23 references3 citations

TL;DR

This paper introduces ESP—a novel method for increasing behavioral complexity in evolved virtual creatures by using a human-designed syllabus of learning tasks, encapsulation to preserve skills, and a pandemonium-inspired mechanism to resolve competing behaviors. The approach enables evolved creatures to achieve approximately double the previous state of the art in behavioral complexity, with two versions: Fast ESP for linear computation growth and General ESP for full morphological adaptation when resources allow.

ABSTRACT

Since their introduction in 1994 (Sims), evolved virtual creatures (EVCs) have employed the coevolution of morphology and control to produce high-impact work in multiple fields, including graphics, evolutionary computation, robotics, and artificial life. However, in contrast to fixed-morphology creatures, there has been no clear increase in the behavioral complexity of EVCs in those two decades. This paper describes a method for moving beyond this limit, making use of high-level human input in the form of a syllabus of intermediate learning tasks--along with mechanisms for preservation, reuse, and combination of previously learned tasks. This method--named ESP for its three components: encapsulation, syllabus, and pandemonium--is presented in two complementary versions: Fast ESP, which constrains later morphological changes to achieve linear growth in computation time as behavioral complexity is added, and General ESP, which allows this restriction to be removed when sufficient computational resources are available. Experiments demonstrate that the ESP method allows evolved virtual creatures to reach new levels of behavioral complexity in the co-evolution of morphology and control, approximately doubling the previous state of the art.

Motivation & Objective

• To address the stagnation in behavioral complexity of evolved virtual creatures (EVCs) over the past two decades, which has largely remained limited to single or few discriminable behaviors.
• To enable the evolution of increasingly complex, multi-behavior repertoires in EVCs where both morphology and control are co-evolved.
• To develop a scalable method that supports open-ended development of complex behaviors without exponential increases in computational cost.
• To allow reuse and combination of previously learned skills through encapsulation, ensuring backward compatibility during evolution.
• To explore the feasibility of evolving rich, diverse, and high-fitness behaviors in virtual creatures through structured, human-guided learning progression.

Proposed method

• The ESP method uses a human-designed syllabus to decompose complex behaviors into a sequence of smaller, intermediate learning tasks.
• Each learned subskill is encapsulated to preserve its functionality and prevent degradation during later evolution, enabling modular reuse.
• A pandemonium-inspired mechanism resolves conflicts between competing drives or behaviors by allowing multiple behavioral modules to 'compete' and select the most appropriate action.
• Fast ESP enforces constraints on morphological changes after the first skill is learned, ensuring computation time grows linearly with added complexity.
• General ESP removes morphological constraints by retesting and reconciling all prior skills after each new skill is added, enabling full morphological adaptation at the cost of higher computation.
• The system supports open-ended evolution by allowing new skills to be built on top of encapsulated, reusable subskills, enabling hierarchical behavior composition.

Experimental results

Research questions

• RQ1 Can the behavioral complexity of evolved virtual creatures be significantly increased beyond the current state of the art, where most EVCs exhibit only one or a few discriminable behaviors?
• RQ2 Is it possible to achieve scalable, open-ended development of complex behaviors in EVCs when both morphology and control are co-evolved, without exponential growth in computational cost?
• RQ3 How can previously learned skills be preserved and reused in evolving creatures that undergo continuous morphological change?
• RQ4 To what extent can a human-designed syllabus of learning tasks guide the evolution of increasingly complex, multi-behavior repertoires in virtual creatures?
• RQ5 Can a hybrid approach combining constrained and unconstrained morphological adaptation enable both efficiency and high-quality, diverse solutions in EVC evolution?

Key findings

• The ESP method successfully increased the behavioral complexity of evolved virtual creatures to approximately double the previous state of the art.
• Fast ESP achieved this increase with only a linear growth in computation time as behavioral complexity was added, enabling scalable evolution.
• General ESP demonstrated greater solution variety and higher fitness by allowing full morphological adaptation across multiple skills, albeit with higher computational cost.
• Encapsulation effectively preserved learned subskills across evolutionary generations, enabling reliable reuse and combination in higher-level behaviors.
• The pandemonium mechanism successfully resolved conflicts between competing behaviors, allowing creatures to dynamically select appropriate actions based on context.
• Experiments confirmed that the method supports open-ended development of complex, multi-behavior repertoires, including advanced behaviors like fight-or-flight and tip-recovery.