Miikkulainen, Risto

computer scientist at UT Austin; associate VP of Evolutionary AI at Cognizant; specializes in evolutionary computation, neuroevolution, cognitive science, and neuroscience

Risto Miikkulainen: Neuroevolution and Evolutionary Computation

Lex Fridman Podcast · Apr 19, 2021

• Risto Miikkulainen
• evolutionary computation
• neuroevolution
• artificial life
• definitions of intelligence
• emotion and consciousness in AI
• theory of mind
• multi-agent simulation
• deep learning architecture search
• reinforcement learning versus evolution
• novelty search
• meaning and mortality

Core ideas

Miikkulainen opens by treating evolution as something that can be studied computationally, not just biologically: running simulated evolution repeatedly, he expects certain capabilities (manipulation, communication, vision, decision-making) to re-emerge reliably because they are useful solutions to recurring problems, while whether something resembling humans specifically would re-emerge remains genuinely open. He proposes a deflationary, functional definition of intelligence — an agent with limited sensory and effector capabilities plus a decision-making mechanism, sufficient to survive or achieve given goals — explicitly independent of biological substrate, and offers a "next level" definition tied to leaving a lasting, useful trace beyond one's own existence (not necessarily unique or Mozart-level, but locally novel). Asked about mortality, he connects the human awareness of finite time to creativity and the drive to build something beyond mere survival, and discusses emotions (including fear) as computational mechanisms for focusing attention and filtering relevant information under time pressure, rather than mysterious extras — while acknowledging deep uncertainty about consciousness itself.

A recurring theme is that intelligence and language are fundamentally social. Drawing on his own research modeling hyenas cooperating to steal kills from lions, Miikkulainen argues complex behavior often emerges from balancing competing "emotion-like" states (fear of the lion, affiliative drive toward the group) coordinated through vocalization and gesture, and that hyenas are a good model system precisely because their cognition is simple enough to observe and quantify (even via measurable neurotransmitter byproducts) compared to humans or baboons. He describes a plausible theory that language originated from exchangeable social roles ("I scratch your back, you scratch mine") providing the cognitive scaffolding for grammar and symbolic reference, and argues social structure — not language per se — is likely the more fundamental substrate, with visual/spatial representation potentially predating language as well, given how much cortex is devoted to vision even during abstract or linguistic reasoning.

On evolutionary computation itself, Miikkulainen frames its core appeal as the ability of algorithms to discover solutions humans wouldn't have thought of, because they lack human biases about what a solution "should" look like. He gives concrete examples: an MIT Media Lab hydroponics project where an evolutionary system discovered basil grows bigger and tastier under continuous light (with no night cycle at all), directly contradicting a constraint the biologists had assumed was necessary; a class assignment where a neural network evolved to make far-away moves in a large tic-tac-toe-like game, causing opponent programs to run out of memory and crash — winning by breaking assumptions rather than out-playing within the rules, which he compares to how Deep Blue's unconventional play unsettled Kasparov. He walks through the basic mechanics of evolutionary computation (representation/genotype-to-phenotype mapping, variation via mutation and crossover, and fitness-based selection), noting that computational evolution relies much more heavily on statistically-guided mutation than biological evolution's crossover-centric approach, and that computation is generally far more impatient (small populations, few generations) than biological evolution's vast timescales and larger, more diverse populations — a difference he suggests may be limiting the creativity of current systems, along with computation's continued inability to replicate biology's "major transitions" (single-cell to multicellular to societal levels of selection).

He distinguishes evolutionary computation from deep learning (suited to tasks with abundant labeled data) and reinforcement learning (individual, lifetime-scale learning suited to tasks with a cost per trial), positioning evolution as advantageous when trials are cheap (e.g., simulation) and exploration of very different, even initially failing, solutions is valuable — illustrated by evolved robots that "fail forward" through faceplants into a genuinely fast run, in contrast to reinforcement learning's tendency toward conservative, stable-but-limited gaits. He extends this to neuroevolution: using evolution to set network weights when no clear supervised target exists, to optimize deep learning architecture choices (hyperparameters, topology, activation and loss functions) that have become too complex for humans to design well by hand, and to a two-timescale model where evolution produces a good "starting point" network that then learns within its lifetime — paralleling biological development, where DNA encodes a starting configuration rather than a finished brain. He is candid about current limits: defining a tractable search space and the cost of training each candidate design remain major bottlenecks.

Miikkulainen discusses several evolved multi-agent behaviors as evidence of emergent sophistication: a simulated predator-prey "arms race" between hyenas and zebras that spontaneously produced group-hunting and split-flight evasion behaviors resembling real gazelle tactics, and a foraging-and-competition simulation in which one robot appeared to deceptively lure a competitor toward far-away food to strand it — behavior he cautiously describes with "anthropomorphized terms" as an emergent, rudimentary theory of mind. He argues a simple, functional theory of mind (predicting an agent's likely discrete states and whether they help or hinder shared goals) may suffice for most human-robot interaction, including autonomous vehicles negotiating with pedestrians, without requiring genuine understanding of mind. On language and vision, he suggests deep integration (rather than one being more "fundamental") is the likely future direction, and that fully understanding either ultimately requires modeling broader world knowledge, goals, and social context.

Asked about honesty and deception in evolved multi-agent systems, Miikkulainen notes that even hyena and bacterial populations reliably produce "cheaters" (individuals who don't participate in costly cooperative risk but still share rewards), and argues societies can tolerate a bounded proportion of cheating and dishonesty while remaining functional — suggesting deliberately engineered diversity (including some rule-breaking variation) in future AI systems could be beneficial rather than purely risky, since fixed, identical, "correct" agents deployed at scale may lack the exploratory diversity from which better solutions emerge. He frames AI as an extension of humanity's tool use and of "meme evolution" (culture and ideas evolving faster than biology), rather than an autonomous entity indifferent to human goals, at least for the foreseeable future, while acknowledging built-in economic and design pressure to keep such systems constrained. On cellular automata and simplicity, he expresses genuine fascination with how minimal rules (DNA's simplicity, Conway's Game of Life) generate unbounded complexity through emergence and self-organization, and highlights "novelty search" — rewarding solutions merely for being different from what's been tried, without a fixed fitness target — as a particularly productive design principle. He closes with life advice modeled on his own field: pursue broad exploration and diversity of experience early (classes, languages, travel), but recognizes exploration eventually gives way to sustained commitment and depth (invoking the "10,000 hours" framing) once a clear calling is found, and frames individual mortality — his own included — as meaningful in the way an individual evolutionary "stepping stone" is meaningful: not always the final solution, but a genuine contribution to a much larger, ongoing process.

Excerpts