Source-linked AI summary
Personalized Recommender Systems for Gym Workouts: A Reinforcement Learning Approach
Roan Rosema, Helma Torkamaan, Masoud Mansoury
TL;DR
Existing workout recommenders often omit prescription details and interaction behavior needed for realistic gym planning. This paper introduces an RL framework that compares exercise-only and full-prescription environments with and without skip-based personalization. PPO is strongest when exercise sequencing and user-specific dosage must be solved jointly, although dynamic interaction and available data impose scope limits.
Problem
Existing workout recommendation approaches often address only exercise selection or subsets of prescription factors, despite practical systems needing sets, repetitions, load, and adaptation to skipping.
Method
The paper models gym sessions as sequential RL problems across four environments combining exercise-only or full-prescription actions with no-skip or skip-enabled interaction.
Results
PPO outperformed all baselines in full-prescription environments, improving over greedy by 0.389 return in FPNS and 0.338 return in FPS.
Takeaways & Limitations
RL offers its greatest advantage when workout recommendation must jointly handle exercise sequencing, dosage, user-specific capacity, and interaction-dependent completion.
Takeaways & Limitations
The evaluation uses synthetic users because public sequential gym data combining prescriptions, context, order, and interaction signals is largely unavailable.
Abstract
from arXiv · showhide
Workout recommender systems aim to help gym users complete effective and engaging training sessions. However, recommending exercises alone is insufficient, as a practical system must also determine appropriate sets, repetitions, and training loads, while adapting to user behavior such as skipping exercises. Existing approaches typically consider only a subset of these factors, limiting their applicability in real-world settings. In this paper, we extend workout recommendation from exercise selection to full workout prescription. We propose a reinforcement learning (RL)-based framework with four environments: exercise-only and full-prescription settings, each with and without skip-based interaction. The full-prescription environments recommend exercises, sets, repetitions, and load, while the skip-enabled environments use user skipping behavior for online personalization. Experiments with synthetic users show that modeling the full prescription task leads to higher rewards and greater user engagement than exercise-only recommendation, highlighting the importance of realistic workout planning in personalized gym recommender systems.
1 Introduction
Practical gym recommendation requires prescribing more than exercises and adapting to user feedback. The paper addresses this gap with an RL framework spanning exercise-only and full-prescription environments, including skip-based personalization.
- Full-prescription recommendations jointly specify exercises, sets, repetitions, load, rest, and progression according to user goals, training status, and capacity.
- Existing workout recommenders typically address only subsets of these interdependent prescription factors, limiting their fit with real gym complexity.
- The proposed RL agent constructs sessions sequentially while prescribing dosage and adapting to feedback such as skipped exercises.
- The framework compares exercise-only and full-prescription environments, with and without skip-based online personalization, against RL and non-RL baselines.
2 Related work
Fitness recommendation extends beyond standard item ranking because activities must fit users’ contexts, goals, abilities, constraints, and behavioral outcomes. Prior work covers several related domains, while this paper adapts RL to gym-based full prescription and explicit skip feedback.
- Fitness recommender systems must account for context, goals, ability, constraints, adherence, safety, motivation, and long-term engagement.
- Prior research spans endurance sports, context-aware and privacy-aware personalization, longitudinal activity recommendation, and sensor-based exercise-form feedback.
- The paper extends a home-fitness RL framework from exercise-only gym recommendations to full prescription with explicit skip-only feedback and online personalization.
3 Our proposed framework
The framework models a single workout as an episodic RL session whose agent selects exercises or full prescriptions step by step. Terminal rewards evaluate session quality, with skip-enabled variants incorporating completion scaling and online interaction.
- Each episode is one eight-step workout session evaluated by exercise order, accumulated workload, prescription quality, and possible skipping.
- The agent operates over an observation-based MDP containing user features, session history, and interaction signals, despite hidden user characteristics affecting skips.
- Exercise-only actions select one exercise, whereas full-prescription actions jointly select an exercise and dosage parameters.
- The four environments differ by exercise-only versus full prescription and no-skip versus skip-enabled interaction, with skip variants using completion scaling and online personalization.
- Full-prescription environments contain 5 set bins, 19 repetition bins, and 21 load bins, producing 432,915 action combinations per step.
- Rewards are terminal and combine sequence qualities such as diversity, routine alignment, and fatigue management, with full-prescription variants adding prescription quality and persistence.
- Skip-enabled rewards scale weighted session quality by the fraction of planned exercises completed, while below-threshold components receive a fixed penalty before weighting.
4 Experiments
The experiments use synthetic users and a simulator to evaluate PPO and non-RL baselines under exercise-prescription and skip-interaction settings. Policies are assessed primarily by mean return, with completion and skipped ratios additionally examined in skip-enabled environments.
- Experimental setup: Missing public sequential gym-recommendation data motivates using structured exercise content, load references, synthetic users, and a controlled simulator.The available data does not jointly provide exercise order, prescription variables, user context, and interaction signals such as skipping or adherence.
- Experimental setup: Skip-enabled environments generate binary skip outcomes from workload, capacity, strain, routine mismatch, recency, step position, and base skip-rate factors.Skip-only feedback updates online estimates of skip bias, capacity, and per-muscle avoidance.
- Experimental setup: Full-prescription environments discretize sets, repetitions, and load into fixed bins over 217 catalog exercises.Load bins span 0.20–1.20 of baseline one-repetition maximum and are converted to working weights using StrengthLevel lookups and the inverted Epley relation.
- Experimental setup: PPO uses the same MLP architecture and hyperparameters across runs, including hidden sizes [256, 256], learning rate 3 · 10^-4, and discount factor γ = 0.99.Policies are evaluated by complete-episode rollouts recording return, reward components, and, where applicable, completion and skipped ratios.
- Baselines: The comparison includes random, greedy, and PSO baselines alongside PPO.Greedy uses reward-aligned sequencing scores and fixed prescription rules, whereas PSO performs offline episode planning with terminal reward evaluation.
- Evaluation: Mean return is the main metric, with 95% bootstrap confidence intervals used to avoid over-interpreting small differences.Skip-enabled environments additionally report completion ratio and skipped ratio.
5 Results
Results favor PPO when recommendation jointly models prescription and skip interaction, while greedy performs best in the simpler exercise-only environments. Skip-based personalization improves adherence awareness but can trade off fatigue or prescription quality, and rapidly reversing user drift reduces reliability.
- Full-prescription results: PPO achieves the highest mean return and is separated from baselines in the full-prescription skip setting.This setting jointly selects exercises, sets, repetitions, and load while adapting to skip feedback.
- Exercise-only results: Greedy performs best in exercise-only environments, random performs worst, and PPO remains competitive with PSO across user-pool regimes.Greedy directly encodes routine alignment, diversity, uniqueness, and strain avoidance, which are also important reward components.
- Full-prescription results: PPO improves over greedy by 0.389 return in FPNS and 0.338 return in FPS in the static pool.It also outperforms PSO by 0.206 in FPNS and 0.198 in FPS; full prescription couples exercise selection with dose selection.
- Skip interaction: Skip-enabled settings reward sessions that are actually completed, whereas no-skip settings reward sessions that look good on paper.Skip-enabled models become more adherence-aware, but optimizing completion can trade off against fatigue or prescription quality.
- User-pool dynamics: PPO reaches similar final performance under static and dynamic user pools, but chaotic, rapidly reversing drift makes skip-based personalization less reliable.This result marks a boundary for the reliability of online personalization under unstable user dynamics.
6 Conclusion
The study extends personalized workout recommendation to gym-based full prescription and finds that RL is most advantageous when jointly handling exercise sequencing, dosage, and user interaction.
- The framework extends workout recommendation from exercise selection to gym-based full prescription, incorporating sets, repetitions, load, skip behavior, online personalization, and evolving user characteristics.
- PPO did not surpass the strong greedy baseline in exercise-only environments, establishing a boundary for RL’s benefit.
- In full-prescription environments, PPO outperformed all baselines including greedy, showing the strongest RL advantage when sequencing and dosage must be solved jointly under capacity constraints.
- Skip-enabled environments produced more adherence-aware policies by optimizing for completed sessions, although completion can tension with other reward components.
- Future work should validate the framework with real users, richer feedback, and longer-term training plans.
GenAI Usage Disclosure
The authors used ChatGPT for editorial assistance but not for experiments, results, data fabrication, or autonomous scientific decisions.
- ChatGPT assisted with shortening, restructuring, grammar, clarity, and spelling, while the authors reviewed and verified the paper’s claims, citations, tables, figures, and final text.