Re-evaluating evaluation

By David Balduzzi, Karl Tuyls, Julien Pérolat, Thore Graepel. 2018

119 citations recorded by the source corpus.

Progress in machine learning is measured by careful evaluation on problems of outstanding common interest. However, the proliferation of benchmark suites and environments, adversarial attacks, and other complications has diluted the basic evaluation model by overwhelming researchers with choices. Deliberate or accidental cherry picking is increasingly likely, and designing well-balanced evaluation suites requires increasing effort. In this paper we take a step back and propose Nash averaging. The approach builds on a detailed analysis of the algebraic structure of evaluation in two basic scenarios: agent-vs-agent and agent-vs-task. The key strength of Nash averaging is that it automatically adapts to redundancies in evaluation data, so that results are not biased by the incorporation of easy tasks or weak agents. Nash averaging thus encourages maximally inclusive evaluation -- since there is no harm (computational cost aside) from including all available tasks and agents.

Subfields: Artificial Intelligence, Evaluation & Benchmarks

How researchers see this paper

The researchers whose trusted networks rank it highest, strongest advocate first. Every researcher gets a different order.

In Artificial Intelligence

In Evaluation & Benchmarks

Citations in the corpus

A Large-scale Study of Representation Learning with the Visual Task Adaptation Benchmark (2019, methodology): “…(i) assigning equal weight to every dataset – there are three data sets that are used in two tasks, and the others are used in one task, (ii) assigning equal weight to each task group: NATURAL , SPECIALIZED , and STRUCTURED , and (iii) the weighting scheme introduced in Balduzzi et al. (2018).”
OpenSpiel: A Framework for Reinforcement Learning in Games (2019, background): “, the agents Rock, Paper, and Scissors in the eponymous game), unlike the Elo rating system [4]; an additional practical benefit is that it is also tractable to compute in general games, unlike ranking systems relying on Nash equilibria [19].”
Emergent Coordination Through Competition (2019, methodology): “Unfortunately, as shown in Balduzzi et al. (2018), in a simple game of rock-paper-scissors, a rock-playing agent will attain high Elo score if we simply introduce more scissor-play agents into a tournament.”
Open-Ended Learning Leads to Generally Capable Agents (2021, background): “For example, averaging will not surface an agent’s failure modes on some tasks if these tasks do not occupy a big part of the task space (Balduzzi et al., 2018).”
Open-ended Learning in Symmetric Zero-sum Games (2019, methodology): “A low-dim representation of the EGS can be constructed via the Schur decomposition, which is the analog of PCA for antisymmetric matrices (Balduzzi et al., 2018b).”
ELF OpenGo: An Analysis and Open Reimplementation of AlphaZero (2019, methodology): ““Nash Averaging 50,000” refers to the Nash averaging rating (Balduzzi et al., 2018), calculated based on a pairwise tour-nament among the same models as in “Selfplay ELO 50,000”.”
Human-Timescale Adaptation in an Open-Ended Task Space (2023, methodology): “A convenient way of using the Pareto improvement criterion to compute a scalar metric is the Nash average method of Balduzzi et al. (2018).”
Differentiable Game Mechanics (2019, background): “An example from Balduzzi et al. (2018b) may help develop intuition for antisymmetric matrices:”
Options as responses: Grounding behavioural hierarchies in multi-agent RL (2019, background): “For example, the game’s intransitive response dynamics (rock beats scissors, which beat paper, which in turn beats rock), may induce endlessly recurring cyclic reinforcement learning challenges [4, 18, 24].”
The Visual Task Adaptation Benchmark (2019, methodology): “…schemes (i) assigning equal weight to every dataset – there are three data sets that are used in two tasks, and the others are used in one task, (ii) assigning equal weight to each task group: NATURAL, SPECIALIZED, and STRUCTURED, and (iii) the weighting scheme introduced in Balduzzi et al. (2018).”
The Benchmark Lottery (2021, background): “[Machado et al., 2018, Bouthillier et al., 2021, Balduzzi et al., 2018, Bouthillier et al., 2019, Musgrave et al., 2020], or “What are the best practices for fair benchmarking?””
Diverse Agents for Ad-Hoc Cooperation in Hanabi (2019, background): “The last issue could be addressed by some form of evaluation that is invariant to redundant agents, such as the use off Nash averaging (in the context of 2-player zerosum games) in [27].”
α-Rank: Multi-Agent Evaluation by Evolution (2019, background): “This rating system, however, comes with two key limitations [64]: first, it has no predictive power in games with intransitive (cyclic) relations in the set of evaluated agents (e.”
Arena: A General Evaluation Platform and Building Toolkit for Multi-Agent Intelligence (2019, methodology): “…based on GUI-conﬁgurable social trees and BMaRSs; (3) the baseline implementations of 5 state-of-the-art multi-agent algorithms for both competitive and collaborative settings; and (4) the sets of benchmark agents/teams for the community to conduct stable and uniform population evaluation [3].”
Toybox: A Suite of Environments for Experimental Evaluation of Deep Reinforcement Learning (2019, background): “Recent investigations into in how the community handles the evaluation and replication of agent performance has exposed some serious challenges that the community needs to address (Henderson et al., 2017; Balduzzi et al., 2018; Clary et al., 2018; Jordan et al., 2018).”
From Motor Control to Team Play in Simulated Humanoid Football (2021, methodology): “Fitness Measure for Football In the multi-agent football task we use Nash Averaging (65) as the fitness measure to drive our PBT mechanism.”
Scaling Scaling Laws with Board Games (2021, background): “This is not necessarily the case, and in fact there are games - such as rock-paper-scissors - where the Elos assigned to each player are entirely uninformative [25]–[27].”