ADAPTER-RL: Adaptation of Any Agent using Reinforcement Learning

Deep Reinforcement Learning (DRL) agents frequently face challenges in adapting to tasks outside their training distribution, including issues with over-fitting, catastrophic forgetting and sample inefficiency. Although the application of adapters has proven effective in supervised learning contexts such as natural language processing and computer vision, their potential within the DRL domain remains largely unexplored. This paper delves into the integration of adapters in reinforcement learning, presenting an innovative adaptation strategy that demonstrates enhanced training efficiency and improvement of the base-agent, experimentally in the nanoRTS environment, a real-time strategy (RTS) game simulation. Our proposed universal approach is not only compatible with pre-trained neural networks but also with rule-based agents, offering a means to integrate human expertise.

Partial advantage estimator for proximal policy optimization

Estimation of value in policy gradient methods is a fundamental problem. Generalized Advantage Estimation (GAE) is an exponentially-weighted estimator of an advantage function similar to $\lambda$-return. It substantially reduces the variance of policy gradient estimates at the expense of bias. In practical applications, a truncated GAE is used due to the incompleteness of the trajectory, which results in a large bias during estimation. To address this challenge, instead of using the entire truncated GAE, we propose to take a part of it when calculating updates, which significantly reduces the bias resulting from the incomplete trajectory. We perform experiments in MuJoCo and $\mu$RTS to investigate the effect of different partial coefficient and sampling lengths. We show that our partial GAE approach yields better empirical results in both environments.

Joint action loss for proximal policy optimization

PPO (Proximal Policy Optimization) is a state-of-the-art policy gradient algorithm that has been successfully applied to complex computer games such as Dota 2 and Honor of Kings. In these environments, an agent makes compound actions consisting of multiple sub-actions. PPO uses clipping to restrict policy updates. Although clipping is simple and effective, it is not efficient in its sample use. For compound actions, most PPO implementations consider the joint probability (density) of sub-actions, which means that if the ratio of a sample (state compound-action pair) exceeds the range, the gradient the sample produces is zero. Instead, for each sub-action we calculate the loss separately, which is less prone to clipping during updates thereby making better use of samples. Further, we propose a multi-action mixed loss that combines joint and separate probabilities. We perform experiments in Gym-$\mu$RTS and MuJoCo. Our hybrid model improves performance by more than 50\% in different MuJoCo environments compared to OpenAI's PPO benchmark results. And in Gym-$\mu$RTS, we find the sub-action loss outperforms the standard PPO approach, especially when the clip range is large. Our findings suggest this method can better balance the use-efficiency and quality of samples.

Visualising Multiplayer Game Spaces

We compare four different `game-spaces' in terms of their usefulness in characterising multi-player tabletop games, with a particular interest in any underlying change to a game's characteristics as the number of players changes. In each case we take a 16-dimensional feature space, and reduce it to a 2-dimensional visualizable landscape. We find that a space obtained from optimization of parameters in Monte Carlo Tree Search (MCTS) is the most directly interpretable to characterise our set of games in terms of the relative importance of imperfect information, adversarial opponents and reward sparsity. These results do not correlate with a space defined using attributes of the game-tree. This dimensionality reduction does not show any general effect as the number of players. We therefore consider the question using the original features to classify the games into two sets; those for which the characteristics of the game changes significantly as the number of players changes, and those for which there is no such effect.

Rinascimento: searching the behaviour space of Splendor

The use of Artificial Intelligence (AI) for play-testing is still on the sidelines of main applications of AI in games compared to performance-oriented game-playing. One of the main purposes of play-testing a game is gathering data on the gameplay, highlighting good and bad features of the design of the game, providing useful insight to the game designers for improving the design. Using AI agents has the potential of speeding the process dramatically. The purpose of this research is to map the behavioural space (BSpace) of a game by using a general method. Using the MAP-Elites algorithm we search the hyperparameter space Rinascimento AI agents and map it to the BSpace defined by several behavioural metrics. This methodology was able to highlight both exemplary and degenerated behaviours in the original game design of Splendor and two variations. In particular, the use of event-value functions has generally shown a remarkable improvement in the coverage of the BSpace compared to agents based on classic score-based reward signals.

Griddly: A platform for AI research in games

In recent years, there have been immense breakthroughs in Game AI research, particularly with Reinforcement Learning (RL). Despite their success, the underlying games are usually implemented with their own preset environments and game mechanics, thus making it difficult for researchers to prototype different game environments. However, testing the RL agents against a variety of game environments is critical for recent effort to study generalization in RL and avoid the problem of overfitting that may otherwise occur. In this paper, we present Griddly as a new platform for Game AI research that provides a unique combination of highly configurable games, different observer types and an efficient C++ core engine. Additionally, we present a series of baseline experiments to study the effect of different observation configurations and generalization ability of RL agents.

AI and Wargaming

Recent progress in Game AI has demonstrated that given enough data from human gameplay, or experience gained via simulations, machines can rival or surpass the most skilled human players in classic games such as Go, or commercial computer games such as Starcraft. We review the current state-of-the-art through the lens of wargaming, and ask firstly what features of wargames distinguish them from the usual AI testbeds, and secondly which recent AI advances are best suited to address these wargame-specific features.

Does it matter how well I know what you're thinking? Opponent Modelling in an RTS game

Opponent Modelling tries to predict the future actions of opponents, and is required to perform well in multi-player games. There is a deep literature on learning an opponent model, but much less on how accurate such models must be to be useful. We investigate the sensitivity of Monte Carlo Tree Search (MCTS) and a Rolling Horizon Evolutionary Algorithm (RHEA) to the accuracy of their modelling of the opponent in a simple Real-Time Strategy game. We find that in this domain RHEA is much more sensitive to the accuracy of an opponent model than MCTS. MCTS generally does better even with an inaccurate model, while this will degrade RHEA's performance. We show that faced with an unknown opponent and a low computational budget it is better not to use any explicit model with RHEA, and to model the opponent's actions within the tree as part of the MCTS algorithm.