October 10th, 2025

October 10th, 2025

Why learn to play games?

Why learn to play games?

Humans are not the only animals to play games; many evolved species, like dolphins, apes, wolves, elephants, as well as countless bird species play games, in their childhood, and later in life

Humans are not the only animals to play games; many evolved species, like dolphins, apes, wolves, elephants, as well as countless bird species play games, in their childhood, and later in life

I often read growing up that the ability to select objectives and to pursue them was a hallmark (if not a definition) of intelligence,

I often read growing up that the ability to select objectives and to pursue them was a hallmark (if not a definition) of intelligence,

This is absolutely not wrong per se, but the ability to play games that life puts in front of us is definitely something that we would associate with cognitive abilities

This is absolutely not wrong per se, but the ability to play games that life puts in front of us is definitely something that we would associate with cognitive abilities

The size of the search space, the intrinsic hardness, the reward sparsity, or the adversarial/cooperative nature of various games may then be used to qualify the type of cognitive abilities required

The size of the search space, the intrinsic hardness, the reward sparsity, or the adversarial/cooperative nature of various games may then be used to qualify the type of cognitive abilities required

Some even consider that life is a game (and seem adamant that they will win it)...

Some even consider that life is a game (and seem adamant that they will win it)...

What are the games of life?

What are the games of life?

That sounds like a famous cellular automaton, or a board game (gifted by American friends) that I used to play as a kid (and that probably told me more about certain facets of American culture than about life or gameplay)

That sounds like a famous cellular automaton, or a board game (gifted by American friends) that I used to play as a kid (and that probably told me more about certain facets of American culture than about life or gameplay)

Ultimately, there is no game of life, there is just life... and then situations that we can describe as games (at least, if we we need to make complex decisions involving time, money, energy, or other resources, and possibly several agents)

Ultimately, there is no game of life, there is just life... and then situations that we can describe as games (at least, if we we need to make complex decisions involving time, money, energy, or other resources, and possibly several agents)

And then, in some sense, we would like to be good at playing the games in front of us... and the thing is that we wish we could prepare for them, but we don't know what these games are in advance... so can we prepare for other games?

And then, in some sense, we would like to be good at playing the games in front of us... and the thing is that we wish we could prepare for them, but we don't know what these games are in advance... so can we prepare for other games?

Transfer Value

Transfer Value

The basic idea about transfer value is that we should be able to assess the _advantage_ brought by learning to play a game in order to play another game

The basic idea about transfer value is that we should be able to assess the advantage brought by learning to play a game in order to play another game

Assuming we learned to play game $G_1$ and we now have to play game $G_2$, how much better are we at that (compared to how we would have)?

Assuming we learned to play game $G_1$ and we now have to play game $G_2$, how much better are we at that (compared to how we would have)?

The idea is that in practice, we will not know $G_2$ in advance, so we will want to be maximally ready for what could come up next, being reasonably agnostic about this

The idea is that in practice, we will not know $G_2$ in advance, so we will want to be maximally ready for what could come up next, being reasonably agnostic about this

From Transfer To Game Benchmarks

From Transfer To Game Benchmarks

As has been discussed many times, games form a platform to evaluate models

As has been discussed many times, games form a platform to evaluate models

But in particular, we could evaluate a variant of a model trained on a game versus that model not trained on that game: this is the essence of the transfer value

But in particular, we could evaluate a variant of a model trained on a game versus that model not trained on that game: this is the essence of the transfer value

Now, this becomes a way to _evaluate games_: the value of a game can be judged in terms of the comparative advantage that a given base model has to learn that game

Now, this becomes a way to evaluate games: the value of a game can be judged in terms of the comparative advantage that a given base model has to learn that game

Of course, this depends on many hyper-parameters (how much one plays the game), on the base model, on the judge model, on the NPC models, on score normalizations, and on many more things, but the idea is that the essence of the game transfer value (i.e. which games help to learn which games) is fairly universal

Of course, this depends on many hyper-parameters (how much one plays the game), on the base model, on the judge model, on the NPC models, on score normalizations, and on many more things, but the idea is that the essence of the game transfer value (i.e. which games help to learn which games) is fairly universal

And hence, games become benchmark for other games, and this opens the path towards navigating that space and discovering new ones

And hence, games become benchmark for other games, and this opens the path towards navigating that space and discovering new ones

In a subsequent post, we will discuss this game navigation and discovery scheme!

In a subsequent post, we will discuss this game navigation and discovery scheme!