Monte Carlo Tree Search: A Review of Drawbacks, Advantages and Limitations in Complex Multiplayer Games

1. Introduction

Monte Carlo Tree Search (MCTS) is a family of search algorithms that builds a decision tree in four stages: selection, expansion, simulation and backpropagation. Instead of requiring a complete evaluation function for every state, the method estimates the quality of actions from numerous simulations, balancing the exploration of rarely visited alternatives with the exploitation of known actions. This characteristic contributed to expressive results in games such as Go, Chess and Shogi, especially in two-player environments with perfect information (BROWNE et al., 2012).

However, contemporary digital and tabletop games present more complex conditions: several opponents, hidden information, randomness, large action spaces and real-time decisions. In these situations, one additional simulation does not always produce a better decision. A larger number of players enlarges the search tree and makes it difficult to model the opponents' responses. Imperfect information can cause strategy fusion when hidden states are handled inadequately, while time pressure reduces the available computational budget. Despite the growing application of MCTS in these scenarios, the literature remains fragmented and concentrated on classic games.

2. Objectives

The general objective of this work was to identify and organise recurring limitations of MCTS in multiplayer digital and tabletop games, especially in environments with imperfect information, stochasticity and time constraints.

As specific objectives, the work sought to gather the strategies that mitigate these limitations and to identify research gaps that can guide future studies and favour more comparable and reproducible evaluations.

3. Methodology

A literature review was carried out on studies published between 2013 and 2025, with searches in the IEEE Xplore, HAL Science Ouverte and Google Scholar databases. The search strategies combined the terms "Monte Carlo Tree Search" and "MCTS" with expressions related to digital games, board games and multiplayer games. Screening considered works whose title or abstract addressed MCTS, a direct variant or a hybrid method applied to competitive or cooperative games. Studies outside the scope, documents without access to the full text and works whose scope was not compatible with the objectives of the review were excluded.

The searches retrieved 24 records, of which two were duplicated between queries, resulting in 22 unique potentially relevant works (17 in English and five in French). After screening, 18 studies were selected for full analysis and four were excluded. The final corpus gathered 13 studies from IEEE Xplore, three from HAL and two found through Google Scholar, and among the selected ones, 15 texts were in English and three in French. The works were classified by game type, algorithm used, environment constraints, reported results and mitigation strategies. The thematic analysis was organised into three structural categories of limitation, performance limits, tactical blind spots and trade-offs in opponent modelling, and two domain-dependent amplifiers, imperfect information and real-time decisions.

4. Results and Discussion

The analysed studies indicate that MCTS performs most consistently in games with well-defined states, a moderate branching factor and a computational budget large enough for a high number of simulations. Among the 18 selected works, eight addressed digital games, six investigated board or tabletop games and the remaining ones analysed card games, general game playing environments or related applications. MCTS-based algorithms were used in 14 studies, often combined with neural networks, action space abstractions, simulation policies and parameter tuning techniques.

The first limitation identified concerns search performance. Although increasing the number of simulations usually improves decision quality, the marginal gains tend to diminish as new simulations start to revisit trajectories that are already sufficiently explored. Results obtained in Go, real-time strategy games and tabletop games show that doubling the computational budget can produce reduced or no benefits after a certain point. In addition, studies on skill depth show that this limit depends on both the complexity of the game and the configuration of the agent. Mitigation strategies include action space abstraction, simplification of simulations and automatic tuning of policies, exploration constants and rollout strategies through the N-Tuple Bandit Evolutionary Algorithm (NTBEA) (GOODMAN; PEREZ-LIEBANA; LUCAS, 2024).

The second limitation corresponds to the blind spots of tactical reasoning. Since MCTS estimates the average value of the sampled continuations, the algorithm can fail in positions that require recognising specific tactical sequences, long-term threats or infrequent configurations. In Go, situations such as ladders and the semeai illustrate scenarios in which a tactical structure that is evident to experienced players may not be identified by the search.

The third limitation involves opponent modelling. In games with three or more participants, the search process must represent not only the decisions of the player at the root of the tree, but also the possible responses of each adversary. Models that keep independent structures for each opponent tend to represent strategic behaviour better, although they significantly increase the computational cost. In contrast, simplified approaches, such as Paranoid Search and generalised policies, reduce this cost but may fail to represent relevant behaviours. The literature suggests that approaches based on selective focus, generalisation of the adversaries' responses and tree decomposition are promising alternatives, although they are still little evaluated in games with a large number of participants (BAIER; KAISERS, 2020).

Imperfect information intensifies these limitations. In card games, Pokémon and other environments with hidden states, determinisation, which consists of assuming one possible configuration of the unknown information and searching as if it were true, can cause the fusion of incompatible strategies. Information Set MCTS, a variation of Monte Carlo tree search designed to handle uncertainty, reduces this problem by grouping states that the player cannot distinguish, preserving only the information actually available at the moment of the decision (COWLING; POWLEY; WHITEHOUSE, 2012). Even so, the randomness inherent to the environment and the need to infer the opponents' behaviour continue to limit the reliability of the estimates produced.

In real-time environments, such as strategy games, MOBAs and first-person shooters, time constraints become a critical factor. MCTS must select actions in very short intervals while managing combinatorial action spaces, multiple units and partial observations. In this context, hierarchical abstractions, macro-actions, action pruning and policies obtained through supervised learning or reinforcement learning appear as recurring strategies (ONTAÑÓN, 2013). In some situations, alternative methods can perform better. The Rolling Horizon Evolutionary Algorithm (RHEA), for example, evolves short action sequences and performs continuous replanning, proving particularly suitable for cooperative games and highly stochastic environments in which the search tree changes quickly (SFIKAS; LIAPIS, 2022).

Applications of MCTS that depend less on real-time decisions were also identified, such as automatic balancing, playtesting and support for game design. In these scenarios, processing time stops being a critical constraint, allowing different configurations to be explored extensively, imbalances to be identified and design alternatives to be evaluated. Although most studies report results favourable to the proposed approaches, some works observed lower performance of MCTS compared with competing techniques, as in the games Kingdomino and Pandemic.

5. Conclusion

The review also identified five relevant opportunities for future research: (i) adoption and standardisation of shared benchmarks for multiplayer games with imperfect information; (ii) systematic evaluation of opponent modelling techniques in games with a larger number of participants; (iii) investigation of the relationship between the complexity perceived by human players and the effective depth of the search; (iv) study of best-of-three and best-of-five competitive series, considering the history of picks and bans; and (v) expansion of MCTS applications in automatic balancing, playtesting and procedural content generation.